Named Entity Recognition (NER) aims to accurately identify entities with predefined semantic categories and clear boundaries from text. In Chinese NER, the absence of explicit word boundaries, the complexity of semantic expressions, and the widespread presence of polyphonic and visually similar characters often lead to semantic ambiguity. Existing methods predominantly rely on character- or word-level information, with insufficient utilization of key linguistic features such as pinyin and radicals, and multi-source heterogeneous feature fusion is typically performed via simple concatenation or weighting strategies, which fail to capture deep semantic correlations among different features and thus limit further performance improvements. To address these issues, this paper proposes a Chinese NER method based on Multi-Feature Hierarchical Fusion (MFHF) to achieve collaborative modeling and deep semantic integration of multi-dimensional linguistic features. Specifically, in the feature representation stage, four types of embeddings—character, pinyin, radical, and lexical—are constructed, where character embeddings are derived from a pre-trained language model to capture contextual semantic information and long-range dependencies, pinyin embeddings encode phonetic sequences to model pronunciation differences and alleviate polyphonic ambiguity, radical embeddings employ a convolutional neural network to model character structures and extract fine-grained semantic features from the glyph level, and lexical embeddings incorporate word-level information via a lexicon matching mechanism to enhance the model’s ability to detect multi-character entity boundaries, thereby improving character representations from phonetic, glyph, and lexical semantic perspectives. To address insufficient interaction and coarse granularity in multi-source feature fusion, a hierarchical cross-attention mechanism is designed, where at the local level, two groups of cross-attention—pinyin–radical and character–lexical—are constructed to model the intrinsic relationships between phonetic and glyph information as well as the structural dependencies between character-level and word-level semantics through bidirectional attention interactions, enabling fine-grained alignment and complementarity among heterogeneous features, and at the global level, the locally enhanced multi-source features are concatenated and further modeled using a multi-head self-attention mechanism to capture long-range dependencies across features, achieving deep semantic integration and generating semantically enriched representations. On this basis, a joint optimization strategy combining multi-task learning and adversarial training is introduced, where auxiliary tasks of pinyin prediction and radical prediction are designed to strengthen feature learning, and gradient-based adversarial perturbations are applied in the embedding space to improve robustness and generalization under complex conditions. Finally, the fused representations are fed into a Bidirectional Long Short-Term Memory (BiLSTM) network for sequence modeling, and a Conditional Random Field (CRF) layer is employed for global decoding to obtain entity recognition results. Experiments conducted on three public Chinese NER datasets, MSRA, Weibo, and Resume, demonstrate that the MFHF model achieves F1 scores of 96.78%, 96.14%, and 71.80%, respectively, outperforming several representative baseline models, with improvements of 1.09, 1.55, and 1.68 percentage points over CPL-NER, GS-Lexicon, and Lattice-LSTM on the respective datasets. In summary, the proposed approach effectively enhances semantic modeling capability and model robustness for Chinese NER through multi-feature hierarchical fusion and joint optimization strategies.

Low-precision training for large language models helps reduce training cost and improve hardware utilization. However, existing high-efficiency low-precision training frameworks mostly rely on native FP8 hardware support, making them difficult to migrate directly to domestic AI accelerators that lack FP8 execution capability. Therefore, a key challenge is how to reconstruct a low-precision training path suitable for domestic accelerators without relying on dedicated FP8 hardware units, while still maintaining training stability and achieving practical end-to-end performance gains. To address this issue, this paper proposes an INT8 dynamic-quantization-based efficient Transformer Engine training scheme for domestic hardware. The proposed scheme redesigns the original FP8 linear-layer computation flow around the integer matrix multiplication capability already available on domestic accelerators, thereby enabling low-precision pretraining of large language models without dedicated FP8 hardware support. In terms of method design, the proposed scheme preserves the dynamic scaling management principle of Transformer Engine and reconstructs the original FP8-dependent linear-layer computation flow into a cross-precision execution path consisting of dynamic quantization, INT8 matrix multiplication, INT32 accumulation, and fused dequantization recovery. This design maps the most computation-intensive matrix multiplication operations onto the underlying integer compute units. To balance implementation feasibility and execution efficiency, a tensorwise dynamic quantization strategy is adopted, in which activations and weights are scaled online according to the dynamic range of each tensor. Combined with the native support of domestic SIMT accelerators for INT8×INT8 integer matrix multiplication and INT32 accumulation, this design enables the domestic adaptation of the core linear-layer operators in Transformer Engine. Furthermore, to address the problems of activation–gradient scale mismatch, quantization error amplification, and convergence degradation that easily arise in numerically sensitive modules such as the input embedding layer and output layer under uniform INT8 quantization, this paper analyzes the numerical characteristics of these layers from the perspectives of gradient propagation and error propagation, and accordingly proposes a hierarchical precision quantization strategy. Specifically, the input embedding layer and output layer remain in BF16 precision to ensure stable gradient propagation and reliable parameter updates; computation-intensive intermediate modules, including attention projection layers and feed-forward networks, adopt an INT8 low-precision path to fully exploit the throughput of integer compute units; scaling factors and some critical intermediate variables are retained in higher precision to balance numerical stability and practical acceleration. On this basis, the proposed scheme is integrated into the Megatron-lm distributed training framework and validated through multi-model pretraining experiments on domestic accelerators. The experiments evaluate Llama2-7B, Llama2-13B, Llama3.1-8B, Qwen3-4B, Qwen3-8B, and Mixtral-8x7B-8L, the last of which is an 8-layer pruned version based on the Mixtral-8x7B architecture. Under the same number of training iterations, the proposed INT8 scheme is compared with the BF16 baseline. The results show that the proposed method maintains training loss curves overall close to those of the BF16 baseline across different models, without obvious oscillation, divergence, or convergence stagnation, indicating that the reconstructed INT8 training path can effectively preserve convergence stability during large-model pretraining. In terms of end-to-end training efficiency, the achieved speedups for Llama2-7B, Llama2-13B, Llama3.1-8B, Qwen3-4B, Qwen3-8B, and Mixtral-8x7B-8L are 1.21, 1.16, 1.17, 1.07, 1.20, and 1.12, respectively, demonstrating stable efficiency gains across models of different scales and architectures. Overall, the proposed method effectively reconstructs the low-precision training path of Transformer Engine on domestic accelerators without native FP8 hardware support. Through the coordinated design of dynamic quantization, an INT8 computation path, and a hierarchical precision quantization strategy, the method achieves stable end-to-end acceleration while maintaining convergence stability. The experimental results indicate that, under current hardware conditions, software-level computation-path reconstruction combined with model-structure-aware precision configuration can effectively unlock the potential of integer compute units, providing a practical solution for efficient pretraining of large language models on domestic platforms.

This study aims to investigate the degradation risks of Generative Artificial Intelligence (GAI) models in self-training loops, with a focus on two core phenomena: content homogenization and the widening divergence between human and machine-generated texts. We select two representative generative models with distinct architectures and build an iterative self-training framework, using the proportion of human data in the training set (α) as the key hyperparameter. Under different initial values of α, we conduct controlled experiments combining two typical dynamic strategies—linear decay and exponential decay—and systematically evaluate the quality, diversity, and human-likeness of generated content using multidimensional performance metrics. The results show that, during self-training, GAI models exhibit a persistent decline in performance, a marked reduction in output diversity, and a gradual increase in the gap between human and machine-generated texts. The linear decay strategy can effectively slow down the decline of information entropy and help maintain content diversity, but it becomes increasingly vulnerable to the cumulative impact of model-generated data pollution in later stages. In contrast, although the exponential decay strategy leads to more pronounced performance fluctuations in the early phase, it achieves superior stability in the long run. Moreover, lightweight unidirectional language models (GPT2) are more prone to falling into a vicious cycle of noise amplification during self-training, whereas bidirectional encoder models (BART), endowed with stronger global modeling capacity, demonstrate greater robustness in the presence of synthetic data contamination. These findings provide important empirical support for optimizing dynamic data-mixing strategies in GAI self-training.

Unmanned aerial vehicle (UAV) power inspection images often contain cluttered backgrounds and variable target scales. These factors limit the image retrieval accuracy. To solve these problems, this paper proposes a power image retrieval network named Swin-FMG. The network is based on frequency-domain coordinate synergy and multi-scale gating. The method uses Swin Transformer as the backbone architecture. First, it proposes a Frequency-domain Coordinate Collaborative Attention (FCCA) mechanism. FCCA combines global spectrum filtering and orthogonal space projection. It effectively suppresses environmental noise and restores the physical continuity of target geometric features. Second, the method designs a Semantic-Guided Multi-Scale Convolutional Gated Fusion (MSCGF) module. MSCGF uses deep semantics to adaptively filter shallow multi-scale textures. It also constructs a dual-stream retrieval representation. This module greatly enhances the perception ability of the model to cope with cross-view scale changes. Finally, the method introduces Low-Rank Adaptation (LoRA) fine-tuning and a joint loss function with hard-sample triplets. These strategies mitigate the overfitting risk on small samples. They also optimize the inter-class separability of the feature metric space. The method is evaluated on a self-built power inspection image retrieval dataset. Experimental results show that the mean Average Precision (mAP) of Swin-FMG reaches 63.15%. The Recall@1 reaches 71.04%. Compared with the baseline Swin Transformer, the mAP of Swin-FMG increases by 4.19%. In conclusion, Swin-FMG effectively strips complex environmental interference and captures scale-invariant features. It significantly improves the image retrieval performance of power equipment while maintaining computational efficiency. The experimental results verify the effectiveness of the proposed method.

In the process of compiler optimization for dynamically typed languages, a large number of runtime check nodes must be inserted due to the uncertainty of runtime types. Existing Redundant Code Elimination (RCE) algorithms based on reachability analysis generally treat all control flow nodes as potentially having side effects. As a result, computations and control flow structures associated with these check nodes are preserved during analysis, making it difficult to safely remove semantically redundant computations and control flows. To address this issue, a semantics-driven RCE method for the Ark runtime is proposed based on a systematic analysis of its compilation workflow and Intermediate Representation (IR) structure. The method begins with observable program semantics, where program behavior is abstracted as a sequence of observable events, including input/output operations, exception throwing, system calls, and termination with specific return values. Based on this abstraction, the RCE problem is formulated as the removal of IR subgraphs under the constraint that observable program semantics remain unchanged. On this basis, a criterion for eliminating runtime check nodes is introduced: a check node and its dependent computations can be safely removed if they produce no side effects and their results are not used by any node that affects observable program behavior. This criterion overcomes the interference introduced by runtime checks and allows the removal of computations that are preserved by traditional approaches despite being semantically redundant. Around this criterion, a semantics-constrained live node propagation mechanism is designed. The process starts with initializing an live node set containing side-effect nodes, followed by expanding this set along data dependency relations. Only nodes that may influence observable program behavior are retained in the set, enabling the identification and elimination of redundant computations and their associated check nodes. Furthermore, to address redundant control flows that cannot be handled by existing methods, control flow graph construction, dominance analysis, and loop structure identification algorithms are incorporated. A method for detecting and eliminating redundant loops and redundant branches is proposed, enabling the overall removal of structures such as empty loops and branches. The proposed method has been integrated into the Ark runtime compilation framework, achieving optimization of redundant computations and control flows at the IR level. Experimental results demonstrate that, in terms of executed instructions, the average reduction across all test cases reaches 3.4%, while representative cases containing redundant control flows achieve an average reduction of 27.4%, with a maximum of 98.26%. In terms of execution time, an average reduction of 3.4% is observed across all test cases, with representative cases achieving an average reduction of 26.4%, and up to 99.99% reduction in loop-intensive programs. Regarding compilation overhead, the execution time of the proposed algorithm accounts for only 2.28% of the total compilation time on average, indicating low additional cost. In overall performance evaluation, the total time of compilation and execution decreases in most representative cases, with a maximum reduction of 94.55%. In addition, validation on 913 runtime unit test cases and 19749 test262 standard test cases shows that no semantic deviation is introduced. Compared with source-level redundancy elimination approaches, further performance gains are achieved at the fine-grained computation and control flow levels, demonstrating the unique advantages of IR-level optimization. The proposed method effectively overcomes the limitations imposed by runtime checks on RCE in dynamically typed languages, significantly improves execution efficiency while preserving semantic equivalence, and maintains low compilation overhead, providing a new approach for compiler optimization in dynamic language runtimes.

Aspect-based sentiment analysis (ABSA) aims to judge the sentiment polarity of specific aspects in texts. Existing methods usually adopt graph neural networks and attention mechanisms to encode the syntactic dependency information and semantic information of sentences. However, only the dependency relations between words can be captured by the syntactic dependency tree, and phrase-level syntactic structure cannot be expressed, thus limiting the model's utilization of phrase-level syntactic information. Moreover, when attention mechanisms are used to capture the semantic features of sentences, they are usually interfered by irrelevant contexts, thus generating semantic noise. Therefore, this paper proposes an aspect-based sentiment analysis model based on syntactic enhancement and semantic denoising. In the syntactic branch, a syntactic constituent tree is introduced to construct a syntactic constituent graph, so as to supplement phrase-level syntactic information. Two types of syntactic information are encoded by the syntactic constituent graph and the syntactic dependency graph respectively, and are dynamically weighted and aggregated through a syntactic fusion mechanism to obtain syntactically enhanced representations. In the semantic branch, a differential attention mechanism is introduced to reduce the attention weights of irrelevant contexts, thereby reducing semantic noise and obtaining denoised semantic representations. In addition, representations fused with external knowledge are obtained by concatenating external knowledge embeddings at the end of word embeddings, so as to help the model better understand sentence semantics. Finally, a multi-feature fusion module is used to fully integrate the three features. Experimental results show that compared with baseline models such as S2GSL, the proposed model improves the accuracy by at least 0.36, 0.83 and 3.13 percentage points on the Laptop, Restaurant and Twitter datasets, respectively, and boosts the F1-score by at least 0.56 and 2.96 percentage points on the Laptop and Twitter datasets, respectively, which verifies the effectiveness of the model’s syntactic enhancement and semantic denoising methods.

Continual anomaly detection focuses on incrementally learning new classes while retaining historical memory. However, the spectral bias and high-frequency artifacts encountered in generative replay severely constrain the fine-grained segmentation of subtle anomalies. To address this, this study proposes DenoiseCAD, a noise-resistant framework based on a cascaded purification architecture, aiming to eliminate feature shifts caused by generative artifacts and prevent the model from capturing spurious features unrelated to defects. First, the study proposes a feature prototype-guided latent space correction mechanism. During the reverse diffusion process, it utilizes the feature prototypes of normal classes as semantic anchors and iteratively rectifies latent variables by calculating feature metric gradients, thereby suppressing distribution shift noise from the source. Second, a task-driven frequency filter is constructed based on parameter sensitivity experiments, implementing a multi-granular spectral joint constraint strategy tailored to data source characteristics to effectively block the propagation of high-frequency artifacts. Finally, anchor-based weight consolidation is implemented. Through isotropic parameter distance constraints, it prevents the model from overfitting to residual noise, thereby establishing a full-pipeline denoising framework from source to terminal. This effectively balances the model's plasticity and stability, successfully alleviates the catastrophic forgetting dilemma, and provides a reliable new paradigm for complex intelligent industrial inspection scenarios. Extensive experiments demonstrate that DenoiseCAD achieves state-of-the-art performance on both the VisA and MVTec datasets. Notably, it yields significant improvements of 2.8% and 1.5% in P-AP over previous state-of-the-art methods.

Unmanned aerial vehicle (UAV) object detection has been playing a crucial role in such fields as intelligent transportation and environmental monitoring. Nevertheless, due to the constraints of multiple factors including target size variations and diverse shooting angles, small object detection in UAV aerial imagery is confronted with prominent problems of drastic scale changes and easy feature attenuation. To tackle the aforementioned issues, this study proposes an improved object detection algorithm for UAV aerial-view scenarios based on YOLOv11n, namely DBD-YOLO. In the feature extraction stage, the DWR multi-scale structure is introduced, which combines dilated convolutions with multiple dilation rates and adaptive channel allocation. This structure can effectively expand the receptive field with low computational overhead and enhance the contextual representation of small objects. In the neck network, a new P2 feature layer is incorporated into the feature fusion process. Bi-directional Feature Pyramid Network (BiFPN) is adopted to realize cross-scale bidirectional weighted fusion, so as to improve the collaboration efficiency between shallow detailed features and deep semantic features. Meanwhile, traditional upsampling is replaced by Dysample point resampling, which not only reduces memory consumption and latency but also maintains fine-grained features. Finally, the DynamicHead, a dynamic adaptive detection head, is introduced. It integrates scale awareness, spatial awareness, and task awareness into a unified framework, and effectively applies the attention mechanism in the object detection head, thereby comprehensively improving the classification and localization performance of small objects in aerial imagery. Experimental results on the VisDrone2019-DET dataset show that the proposed DBD-YOLO algorithm achieves 45.2% in mAP50 and 27.4% in mAP50-95, representing an increase of 12.1% and 8.1% compared with the baseline , respectively. At the same time, the number of model parameters remains roughly at the same level as the baseline, realizing a dual breakthrough in both detection accuracy and computational efficiency.

To address the challenges of large-scale variations in student behaviors, dense target distributions, and insufficient recognition accuracy for back-row students in classroom scenarios, this paper proposes an improved student behavior recognition algorithm, termed MSD-YOLO, based on the YOLO11n baseline model. First, a Multi-Scale Behavior Perception module is introduced into the backbone to enhance the network’s ability to perceive behavior features at different scales, thereby alleviating the scale inconsistency between front-row and back-row students during the feature extraction stage. Second, a Semantic–Spatial Deep Fusion module is designed in the neck to strengthen the interaction between high-level semantic information and low-level spatial details, improving the discriminative representation of features in dense classroom scenes. Finally, a Dual-Scale Context Aggregation module is embedded before each detection head. By integrating global contextual information with a feature re-calibration mechanism, the proposed module further enhances the network’s ability to distinguish small-scale student behaviors, thereby improving the recognition accuracy of back-row students during the detection stage. Experimental results demonstrate that, compared with the YOLO11n baseline, MSD-YOLO achieves improvements of 3.2% and 3.7% in mAP@0.5 and mAP@0.5:0.95, respectively, on the self-constructed dataset. On the public STBD-08 dataset, the corresponding improvements reach 2.4% and 2.6%. Moreover, with only a slight increase in computational cost and model parameters, the proposed method maintains favorable real-time performance, validating its effectiveness and practical value for classroom student behavior recognition tasks

Online service reputation measurement aggregates user feedback to generate service reputation, which helps users judge service credibility in the absence of sufficient information. However, due to the dynamic evolution of the service environment, service quality, user quantity and user preferences keep changing over time. Reputation measurement methods that only focus on a single time point cannot reflect these changes timely and accurately. In addition, service reputation measurement mechanisms that fail to consider the maximization of user group satisfaction are difficult to attract users to give evaluations consistent with their real experience. This leads to some services being assigned false reputation values. To address these issues, this paper proposes an online service reputation measurement method for maximizing user group satisfaction. First, this paper models online service reputation measurement in dynamic environments as a Partially Observable Markov Decision Process (POMDP) optimization problem for maximizing user group satisfaction. Second, aiming at the inconsistency of user group evaluation criteria, this paper adopts large language models to calculate the reward function and measure user group satisfaction accordingly. Finally, this paper uses the Rainbow DQN algorithm to solve the optimization problem. Experiments are conducted on two public datasets, namely Movielens and Yelp, and multiple large language models are used for evaluation. Results show that the proposed method can generate reputation measurement results consistent with the preferences of most users, thus achieving the maximization of user group satisfaction and verifying the effectiveness of the method.

Stochastic Configuration Networks (SCNs) incorporate randomized learning mechanisms into neural network training to enhance modeling efficiency and employ a data-dependent supervisory mechanism to ensure the universal approximation capability of the model. However, during the incremental construction process, the computation of hidden layer output weights after each newly added hidden node relies on repeated calculation of the pseudoinverse of the hidden layer output matrix, which limits the training efficiency to some extent. In addition, while randomized learning improves modeling efficiency, it inevitably introduces potential redundant hidden nodes and parameters. To address these issues, this paper proposes a group-sparse learning method for Incremental Regularized Stochastic Configuration Networks (GSL-IRSCN). First, to improve the training efficiency of regularized SCNs in the incremental modeling process, an incremental output-weight updating strategy for L2-regularized SCNs is developed based on the Woodbury matrix identity, thereby avoiding repeated computation of the inverse of the regularized normal matrix and effectively reducing the computational cost of the model. Then, to mitigate redundancy in hidden nodes induced by the randomized learning mechanism, group L1/2 regularization is introduced and optimized via the Alternating Direction Method of Multipliers (ADMM), achieving sparsity constraints on redundant nodes and simplifying the network architecture. Experimental results on four benchmark datasets from UCI and KEEL demonstrate that the proposed GSL-IRSCN outperforms existing comparative methods in terms of both training efficiency and model compactness.

参考伪装目标检测（Ref-COD）旨在依托参考图像或文本，精准分割指定伪装目标，是伪装目标检测领域的新型任务。大部分现有方法仅采用单一模态参考信息，在多源参考信息融合及跨模态特征适配方面存在明显局限，难以充分发挥参考指导价值。为此，本文提出一种基于文本-图像多模态融合的Ref-COD网络（TIFNet），实现多源信息高效利用与精细检测。首先，通过金字塔视觉Transformer（PVT）编码器、冻结显著目标检测（SOD）编码器及对比语言-图像预训练（CLIP）编码器，分别提取输入图像、参考图像及参考文本的多阶段特征；设计多键值参考融合模块（MRFM），完成跨模态特征对齐与深度融合，强化参考信息定向指导作用；引入参考空间通道增强模块（RSCM），从双维度实现融合特征与参考特征的双向互增强，消解模态差异；最后利用参考自适应归一化模块（RANM），聚焦关键像素细节，提升模型对多样化伪装场景的自适应能力。大量实验结果表明，该方法相较于近年来主流最优（SOTA）方法，已在R2C7K数据集上的 、 、 、 评价指标上分别取得了0.869、0.929、0.786、0.022的结果，展现出了显著的优势，有效提升了复杂场景下指定伪装目标的分割精度与鲁棒性，为多源信息驱动的伪装目标检测提供了新思路。

To address challenges in thyroid ultrasound nodule segmentation, including blurred boundaries, low contrast, and highly variable small lesions, this paper proposes an improved model named MAD-UNet. The model improves contour delineation by strengthening cross-layer feature transfer consistency and deformable context modeling. A Multi-Directional Separable Attention Module (MDSAM) is embedded in the skip connections between the encoder and the decoder. MDSAM applies direction-aware channel–spatial joint attention to reweight key edge responses. This design enhances the consistency between shallow spatial details and deep semantic information. It strengthens boundary localization and alleviates gradient attenuation during deep network training. In addition, the Transformer encoder depth is extended to 24 layers to better model long-range dependencies and global context. Furthermore, a Deformable Adaptive Multi-Scale Context Module (DAMCM) is introduced. DAMCM combines deformable modeling with multi-scale context aggregation. It enables adaptive fusion of local structure alignment and global context supplementation. It improves representation of irregular contours and small targets. On the TN3K, DDTI, and Shanghai Sixth People's Hospital THN-L datasets, the Dice scores reach 89.10%、90.53% and 91.17%, respectively. The overall performance exceeds the TransUNet baseline on all datasets. Complexity evaluation shows 215.27M parameters, 65.96G FLOPs, and an inference speed of 111 FPS. Visualization analysis shows stronger robustness for nodule contours under complex ultrasound conditions. The experimental results verify the effectiveness of the model in fine boundary delineation and small lesion recognition. The method provides a basis for subsequent deployment and optimization in clinical application scenarios.

To address the issues of poor adaptability of static strategies, strategy space explosion, and performance jitter in collective communication on domestic GPGPU platforms, this paper proposes an offline automatic tuning, communication strategy optimization, and consolidation method for domestic heterogeneous computing platforms. The proposed method constructs a multidimensional performance space model over communication primitives, message sizes, and node scales, and obtains performance data through systematic offline benchmarking. To mitigate the impact of system noise in heterogeneous environments, a filtering mechanism based on default strategy comparison and significance thresholding is designed. Specifically, the default strategy is first used as a baseline to evaluate performance differences, and statistical analysis is then applied to identify communication strategy combinations with significant performance advantages, thereby enabling communication strategy optimization.Furthermore, an interval-based strategy model is constructed to map discrete sampling points into continuous message size ranges, and the optimized strategy mapping is embedded into the internal decision logic of the RCCL communication library. Experimental results on domestic heterogeneous clusters demonstrate that the proposed method enables automatic and accurate strategy selection without introducing any additional runtime overhead. Compared with default strategies, the average bandwidth utilization of Reduce and AllReduce operations is improved by 22.4% and 24%, respectively. By leveraging offline tuning and strategy consolidation, the proposed approach effectively avoids the overhead and instability caused by dynamic search, and provides an efficient and practical solution for improving communication efficiency and system stability in large-scale distributed training systems.

To address the problems of generator training confusion, insufficient image detail restoration and incomplete haze removal in existing unsupervised image dehazing methods based on CycleGAN, an unsupervised image dehazing network based on High-frequency Information Enhancement (HIE-Net) is proposed. First, a Multi-Branch Dehazing Network (MBDN) is constructed. The network implements unified encoding of the image feature space through a shared encoding module, and adopts a multi-branch decoding module to achieve differentiated adaptation and precise decoding for features corresponding to different haze densities. Meanwhile, unsupervised constraints are established based on the Atmospheric Scattering Model (ASM) to regularize the training process of the generator. Secondly, a High-Frequency Multi-scale Enhancement Module (HMEM) is designed. A bidirectional guidance mechanism is built based on a large-kernel grouped attention gate. Through the bidirectional interaction between hazy region features and enhanced high-frequency information, the module synchronously achieves the multi-scale enhancement of both hazy region features and high-frequency information including image textures and edges. Finally, a Channel Feature Purification Module (CFPM) is introduced. A channel cross-attention mechanism is adopted to accurately screen haze-sensitive channels, suppress the interference caused by haze residuals in the feature fusion stage, and optimize the allocation of channel feature space. Meanwhile, a spatial cross-attention mechanism is leveraged to capture the haze density correlation and spatial dependency across different regions, thus achieving the fine-grained purification of deep features. Experimental results demonstrate that on the BeDDE dataset, HIE-Net achieves 21.20 dB, 0.779, and 0.198 in PSNR, SSIM, and LPIPS, respectively, which provides a novel insight for the field of image dehazing.

In traffic monitoring and public security scenarios, relying solely on ground-view or aerial-view vehicle re-identification often fails to meet the requirements of large-scale, complex, and multi-scene perception. Ground-view images contain rich visual details but suffer from limited field of view and frequent occlusions, whereas aerial views offer wide-area coverage but usually depict vehicles with small sizes and insufficient details, leading to degraded recognition performance. Therefore, fusing ground and aerial viewpoints for cross-view vehicle ReID has become a key research direction for enhancing large-scale traffic perception. However, this task is confronted with several challenges, including severe scale variations, large cross-view appearance discrepancies, intra-class distances exceeding inter-class distances, and limited cross-scene data. To this end, we propose a large model-based semantic enhancement method for cross-view vehicle re-identification. Built upon the CLIP-ReID multimodal framework, the proposed approach first employs Qwen-VL-Plus multimodal large model to automatically generate fine-grained structured descriptions for vehicle images, and then leverages Qwen-Max language large model to fuse semantic information from ground and aerial viewpoints, yielding a unified and stable cross-view semantic representation. This representation is further injected into a two-stage image–text contrastive learning scheme to strengthen the model’s domain generalization ability under cross-scene and cross-platform conditions. To promote practical deployment and subsequent research, we also construct a cross-view ground–aerial vehicle image dataset covering multiple flight altitudes, acquisition devices, and scene conditions, and design domain-generalization-oriented data splits and evaluation protocols as a new benchmark. Experimental results demonstrate that the proposed method significantly outperforms pure visual baselines on multiple metrics and achieves superior performance to state-of-the-art algorithms in cross-scene domain generalization tests, validating the effectiveness of semantic enhancement for cross-view vehicle re-identification. The proposed method shows strong application potential and engineering value in intelligent traffic surveillance, UAV-based patrol, and regional security.

As the scale and complexity of complex intelligent systems represented by high-performance computing systems and embedded systems continue to grow, automated anomaly detection of logs, as core operational data, has become critical to ensuring reliable system operation. Traditional log anomaly detection methods driven by machine learning and deep learning focus mostly on log sequence modeling, and suffer from insufficient semantic understanding and limited generalization ability. Large language models have effectively overcome this limitation with their outstanding semantic understanding and contextual reasoning capabilities. Since the rise of large language model technology, relevant research has emerged rapidly, but achievements are scattered across multiple technical paths and lack a systematic review. This paper provides a comprehensive survey of log anomaly detection methods based on large language models, selects 35 core literatures, and establishes a unified technical classification framework. Existing methods are categorized into five technical routes: prompt engineering, retrieval-augmented generation, domain fine-tuning, reinforcement learning, and large-small model collaboration. Research shows that supervised fine-tuning is the most widely used technical route at present, while the large-small model collaborative architecture, as an emerging paradigm, is shifting the research focus from pure pursuit of detection accuracy to balancing inference efficiency and industrial deployability. Current evaluation systems focus heavily on detection performance metrics, with insufficient attention to efficiency overhead and interpretability. Finally, this paper identifies the inference latency bottlenecks and data privacy challenges of large language models in processing ultra-long massive log streams, and proposes insights into frontier directions such as lightweight deployment and online continuous learning.

Causal relationship mining aims to reveal latent causal mechanisms from complex data. Existing studies, mostly based on Bayesian network frameworks or simple filtration of association rules, generally face challenges such as low mining efficiency and difficulty in controlling unobserved confounding variables, resulting in insufficient accuracy and robustness of causal identification. To address this, this paper proposes a fast causal rule mining algorithm. This algorithm utilizes a prefix-tree structure for frequent pattern mining and integrates multiple pruning strategies to significantly enhance mining efficiency. Furthermore, it introduces a covariate mechanism and a matched transaction pair technique to effectively control confounding factors, thereby improving the reliability of causal rules. Experimental results demonstrate that the computational efficiency of the proposed algorithm is improved by 3 to 4 orders of magnitude compared to baseline algorithms. On large-scale datasets, its execution time is further reduced by 30%–50% compared to similar variants. In terms of accuracy, compared with baseline causal methods, the proposed algorithm maintains a stable Precision in the range of 0.69–0.90 and generally achieves an improvement of over 40%–60% in F1-score. These results fully validate the efficiency and superiority of the proposed algorithm in large-scale causal rule mining tasks.

Problem understanding is a critical prerequisite for achieving automated geometric theorem proving. However, existing approaches commonly suffer from excessive reliance on feature engineering and limited generalization capabilities, making them inadequate for effectively supporting automated problem solving. To address this challenge, this paper proposes a large language model-based method for geometric problem understanding by fine-tuning the Qwen2.5 base model and integrating chain-of-thought reasoning with k-nearest neighbor (KNN) retrieval-augmented generation. Furthermore, to enhance the accuracy of semantic translation, we introduce an agent-based hallucination detection and correction mechanism, which significantly mitigates hallucination issues during problem understanding. Experimental results demonstrate that the proposed method achieves an accuracy of 88.85% and a recall of 89.12% on the intent understanding task of the self-constructed dataset, significantly outperforming the baseline model. On the Geometry3K dataset, it attains an accuracy of 94.86% and a recall of 94.18%, exhibiting superior performance compared to the Inter-GPS method. Additionally, comprehensive ablation studies and comparative analyses under various parameter configurations further validate the superior performance and adaptability of our multi-strategy hybrid approach.

The core issue in underwater optical imaging lies in the scattering effects of water, particularly backscattering, which creates an approximately uniform haze-like background during image formation, severely obscuring the structural information of targets. This significantly limits the effective application of underwater vision systems in high-turbidity environments. To address this challenge, this paper proposes an underwater imaging framework that deeply integrates physical processes with computational imaging methods. The core concept is to transform the originally difficult-to-model problem of global strong scattering into a locally separable problem with well-defined geometric and statistical characteristics, through physical scanning and light field redundancy constraints. In terms of implementation, the framework first decomposes wide-area scattering into localized scattering across sequential frames using line-structured light scanning. Subsequently, virtual aperture technology is employed to preprocess light field data based on structured light geometric priors, thereby constraining scattering regions. Furthermore, the angular redundancy of the light field is utilized to construct Epipolar Plane Images (EPI), and low-rank decomposition is applied to separate the backscattering component, which exhibits low-rank properties, from the target signal, which exhibits sparsity. Finally, high-quality underwater images are obtained through sequential frame stitching and luminance homogenization. System experiments were conducted within a turbidity range of 10–30 NTU (Nephelometric Turbidity Units). The results demonstrate that the proposed method consistently outperforms comparative approaches under various turbidity conditions, showing stable improvements in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and no-reference image quality assessment metrics. Particularly under high-turbidity conditions, the method exhibits stronger robustness, with significantly less degradation in imaging quality as turbidity increases compared to other methods. This validates the effectiveness of the physics-computation collaborative imaging framework in complex scattering environments.

With the widespread use of digital images in social media, they have become the core carrier of information dissemination. The rise of powerful and easy-to-use image editing software and generative artificial intelligence technology has lowered the threshold for creation while also providing a more covert way for malicious image tampering, leading to the accelerated spread of false information. Tampering will leave specific tampering features in the image, which constitutes the core basis of image tampering detection technology. In the face of increasingly complex and diverse tampering methods, most of the existing reviews focus on a single technical route, and lack systematic comparison and integrated analysis of image tampering detection techniques. To this end, this paper constructs a three-dimensional classification system of "feature traceability-extraction method-detection task", summarizes the image tamper detection technology into two categories based on manual features and deep learning features, and carries out the following work: First, the system reconstructs the classification framework of image tamper detection technology based on manual features, and integrates the scattered manual features in traditional research into three categories: camera system features, pixel-level features and format-related features. The physical mechanism and improvement effect of the performance optimization strategies of 14 typical image tamper detection technologies are deeply analyzed, and the shortcomings of the existing review in the systematic analysis of image tamper detection technology based on manual features are thoroughly analyzed. Second, the image tamper detection technology based on deep learning is sorted out in architecture, and the generative image tamper detection technology is analyzed. Third, the composition, characteristics and limitations of the existing tampered image datasets are summarized and commented on, and the selection of datasets is provided with a selectable basis. Finally, this paper summarizes and looks forward to the future research direction and development trend of this field, and points out some key scientific issues that need to be solved urgently, in order to provide reference for subsequent research.

To address the issues of identity and audio consistency in existing talking face generation models, a Transformer-based diffusion for talking face generation is proposed. First, to improve identity consistency, a global-local collaborative identity alignment module is designed. This module utilizes attention pooling to aggregate global identity representations and introduces a learnable positional encoding matrix to accurately capture local facial geometry, thus significantly enhancing the ability to preserve identity information. Second, to improve audio consistency, a multi-level feature staggered fusion method based on a diffusion Transformer is proposed. Audio and identity features are deeply fused in each Transformer layer, and a multi-stage training strategy is combined to make the generated lip movements more natural. Experimental results on the public datasets LRS3 and HDTF show that, compared with existing methods, the proposed model achieves better performance in terms of the Sync-C and CSIM metrics.

In multi-label fuzzy data, feature redundancy, complex interaction relationships between features, and unequal feature contributions are commonly present, which affect the classification performance of multi-label learning. To address these issues, ReliefF-β algorithm is proposed to assign feature weights, and a multi-label feature selection method based on feature weighted interaction is presented. Firstly, feature similarity and label similarity are constructed for multi-label fuzzy data. A regulating parameter β is introduced to fuse the two similarities and construct a global sample similarity, then ReliefF-β algorithm is proposed for feature weighting. On this basis, multi-label weighted fuzzy rough set is introduced based on feature weights, and uncertainty measures such as weighted fuzzy entropy and weighted fuzzy mutual information are defined. The related properties and relationships among these measures are studied. Furthermore, a feature weighted evaluation function is defined by considering feature relevance, redundancy, and interaction, then a multi-label feature selection algorithm based on feature weighted interaction is proposed. Finally, comparative experiments are conducted under two classifiers. The results show that, compared with other comparison algorithms, under ML-KNN, the proposed method improves Average Precision (AP) by 8.79% on average, while Hamming Loss (HL), Ranking Loss (RL), Coverage (CV), and One-Error (OE) are reduced by 5.06%, 15.33%, 10.97% and 23.06%, respectively. Under BRDT, AP is improved by 4.06%, and HL, RL, CV, and OE are reduced by 8.60%, 10.28%, 7.19% and 5.89%, respectively. Ablation studies and statistical tests further verify the effectiveness of the proposed method.

Small object detection in complex scenarios has long grappled with two major technical bottlenecks: the propensity for weak object features to attenuate within deep neural networks, and the severe interference caused by environmental background noise. To address these challenges, this study proposes WF-DETR, an end-to-end real-time small object detection model. In the feature extraction stage, a Feature Weaving Network (WeaveNet) is designed. Diverging from simple hierarchical stacking, WeaveNet employs a heterogeneous feature weaving strategy. Leveraging a cross-level feature mutual correction mechanism, it tightly interweaves and bidirectionally calibrates deep semantic information with shallow geometric details. This approach effectively suppresses the attenuation of spatial information during feature transmission and mitigates small object feature loss, all while maintaining high-level semantic strength. Inspired by human visual physiological mechanisms, the neck network incorporates a FoveaFormer module. By simulating the human foveal imaging mechanism via adaptive sparse attention and gating units, this module dynamically filters redundant background noise and focuses on high-value target regions, significantly enhancing feature purity. Furthermore, a Haar Wavelet Downsample (HWD) operator is introduced to reconstruct the downsampling process. From a frequency domain perspective, this overcomes the irreversible loss of high-frequency texture details caused by traditional pooling, further augmenting the discriminability of small object features. Experimental results on the VisDrone2019 benchmark dataset demonstrate that the proposed model achieves mAP@0.5:0.95 of 23.7% and an inference speed of 166.3 FPS. These results fully validate the real-time performance and superiority of WF-DETR in small object detection tasks within complex backgrounds.

Cloud computing offers efficient data storage and management, facilitating convenient data sharing and access. However, ensuring data security and user privacy in open cloud environments remains a critical challenge. Ciphertext-policy attribute-based encryption (CP-ABE) has been widely adopted to enforce fine-grained access control over data stored in cloud servers. Nevertheless, existing schemes still face limitations in handling hierarchical data and tracing malicious ciphertexts, making it difficult to simultaneously achieve efficient multilevel access and data provenance assurance. To address these challenges, this paper proposes a hierarchical attribute-based access control scheme with traceability over ciphertext. First, a hierarchical CP-ABE framework is employed to construct an efficient multilevel access mechanism. By integrating multiple hierarchical access trees into a unified structure, the scheme enables encryption and decryption of data at different levels under a single policy, significantly reducing computational overhead. Second, a zero-knowledge proof-based signature mechanism is introduced to securely bind ciphertexts with their creators while preserving data owner anonymity, enabling accurate tracing of malicious ciphertext sources. Finally, security analysis demonstrates that the proposed scheme can effectively resist chosen-plaintext attacks. Experimental evaluation shows that, compared with existing approaches, the scheme achieves lower encryption and decryption overhead, making it well suited for secure, efficient, and traceable data sharing in cloud environments.

Micro-expressions are fleeting, involuntary facial muscle movements that can reveal genuine emotions individuals attempt to conceal. However, micro-expression recognition faces numerous challenges, including short duration, low intensity, prominent local features, limited scale of public datasets, and significant individual differences, which constrain the recognition accuracy and generalization capability of traditional methods. To address these issues, this study proposes a single-stream fine-grained micro-expression recognition method based on dynamic routing experts. Inspired by the mixture-of-experts model, this method replaces the traditional multi-head self-attention layer in Transformers with a dynamic routing expert mechanism. It dynamically selects expert networks through a sparse activation strategy and leverages a collaboration mechanism among experts to enhance feature representation capability, thereby improving model representational capacity while maintaining computational efficiency. Additionally, a multi-grained asymmetric aggregation module is designed, which integrates orientation-aware convolution and channel attention to effectively decouple spatial features and adaptively adjust feature granularity at different network levels, enabling more precise capture of subtle directional movements and local texture variations in micro-expressions. Experiments conducted on three public datasets, SAMM, SMIC, and CASME II, demonstrate that the proposed method significantly outperforms mainstream approaches. On the composite dataset, the method achieves an unweighted average recall of 87.65% and an unweighted F1-score of 87.21%. The experimental results validate the effectiveness of this method in capturing subtle dynamic features of micro-expressions, providing reliable technical support for emotion recognition in complex scenarios.

The widespread application of robots and vision systems in factories has promoted mixed-line production characterized by small batches and multiple product varieties, while also sharply increasing the diversity of target size specifications and the uncertainty of arrival sequences, thereby making stacking tasks at many transition stages of production lines still highly challenging. With the increase in the number of targets in the sequence, it becomes difficult to guarantee both the solution time and the solution quality of the stacking task. To address the above issues, a hybrid optimization algorithm with stimulus memory for sequence stacking tasks is proposed. The algorithm decomposes the sequence stacking task into two subtasks: combinational block knowledge base construction and stacking decision optimization. First, basic target combinations satisfying quality thresholds are searched for in the initial sequence of targets to be stacked, so as to construct a combinational block knowledge base. During this process, a stimulus memory mechanism is introduced to dynamically update the existing combinational knowledge. Subsequently, each combinational block is equivalently treated as a macro-target, and the placement sequence and placement orientation of all targets are jointly optimized. Comparative experimental results on datasets with different size distributions show that, compared with the baseline algorithms, the proposed algorithm can reduce the solution time of stacking plans by at least 4.94% while achieving the optimal average filling rate of stacking space, which verifies its effectiveness in sequence stacking tasks. The ablation experimental results show that the proposed complete algorithm achieves the best performance in terms of solution time, which validates the rationality of the proposed algorithmic architecture.

To address the problem that, in radar emitter individual identification, a single continuous-pulse model cannot simultaneously capture global temporal information and fine-grained single-pulse features, while a single-pulse model lacks global dynamic information, thereby limiting recognition performance in complex electromagnetic environments, this paper proposes a dual-branch lightweight fusion recognition method. First, the original pulse sequence is segmented into two types of data, namely continuous pulse sequences and single pulses, through continuous pulse segmentation. Corresponding datasets are constructed for the inter-pulse sequence branch and the single-pulse branch, and a continuous-sequence model and a single-pulse model are trained separately to extract inter-pulse temporal features and fine-grained intra-pulse features, thus enabling complementary modeling of the two types of information. Subsequently, two fusion strategies, namely feature-level fusion and decision-level fusion, are designed. In the feature-level fusion strategy, a gating mechanism is introduced to learn the importance weights of features from different branches, so that continuous-pulse features and single-pulse features can be adaptively weighted to construct a joint feature representation. In the decision-level fusion strategy, the probability outputs of the two models are integrated by soft voting to improve recognition stability. To verify the effectiveness of the proposed method, comparative experiments and ablation studies are conducted on a measured radar dataset. The results show that both fusion strategies outperform the individual models. Specifically, decision-level fusion improves the recognition accuracy by approximately 8 percentage points over the single continuous-pulse model and by about 3 percentage points over the single single-pulse model. Moreover, feature-level fusion achieves the best recognition performance while reducing the number of model parameters by two orders of magnitude compared with the baseline model. The results demonstrate that the proposed method can maintain high recognition accuracy while also exhibiting favorable lightweight characteristics and strong potential for engineering applications.

In recent years, although self-supervised skeleton-based action recognition has made progress, it still faces two types of training bias under strong augmentations: imbalance in local perturbation allocation can easily lead to over-perturbation of critical motion segments and insufficient variation in low-dynamic regions; in multi-positive contrastive learning, non-target positive samples participate in normalization competition, which can easily cause target conflicts and weaken representation aggregation. To address this issue, this paper proposes DCD-CLR, a self-supervised contrastive learning framework for collaborative optimization of view construction and objective construction, namely Dual-end Collaborative Debiasing Contrastive Representation Learning, to improve the quality of skeleton representation learning from the two aspects of augmentation allocation and contrastive objective. On the view side, Continuous Dynamic Saliency Augmentation (CDSA) is designed to integrate frame-difference energy and data-level joint motion priors, construct a frame-joint dynamic intensity map, and perform continuous, region-level, and sample-adaptive scheduling of spatiotemporal perturbation magnitudes, thereby improving view diversity while preserving critical motion segments. On the objective side, Target-Isolated InfoNCE (TI-InfoNCE) is proposed as a target-isolated debiased multi-positive contrastive objective, which removes the remaining positive samples when computing the normalization term of the target positive sample, so as to reduce competition interference among positive samples and improve the boundary clarity of the representation distribution. Under the linear evaluation setting, the proposed method achieves recognition accuracies of 85.9%, 79.6%, and 92.6% on NTU60 xsub, NTU120 xset, and PKU-MMD I, respectively; combined with the results of representation distribution visualization, transfer evaluation, and noise interference experiments, it is shown that the proposed method has good stability, generalization ability, and robustness.

With cyberattacks becoming increasingly sophisticated and covert, improving the representation and recognition of complex traffic patterns has become an important issue in intrusion detection. Although existing methods have improved detection performance, directly modeling complex network traffic still suffers from insufficient feature representation. To enhance local correlations and structural information among features, many studies transform one-dimensional traffic features into two-dimensional image-like representations for deep feature learning. However, limited by feature dimensionality and encoding schemes, such traffic images are usually small and structurally constrained, making fixed enhancement strategies insufficient for capturing differences among attack patterns. Meanwhile, class imbalance further restricts the recognition of minority attack classes. To address these issues, this paper proposes a network intrusion detection method based on dynamic selective feature enhancement. At the representation level, a multi-scale feature enhancement module adaptively fuses features with different receptive fields to alleviate the representation limitations of small traffic images. At the decision level, a dynamic adaptive module combined with minority-class attention selectively strengthens key responses to improve minority-class recognition. Experimental results show that the proposed method achieves 96.49% accuracy, 95.11% precision, 96.32% recall, and 95.50% F1-score on NSL-KDD. It also maintains good detection performance on UNSW-NB15 and shows stable performance in a simulated streaming environment built on TON-IoT-Network.

In recent years, RGB-based hand mesh reconstruction has attracted extensive attention. Existing methods mainly rely on stacking complex visual modules to improve reconstruction accuracy, but this often incurs high computational cost and makes it difficult to satisfy the requirements of real-time applications. To address this issue, this paper introduces natural language information during training, injecting high-level prior knowledge into the network to enhance visual feature representation. Since the text branch is used only for supervision during training, it does not increase the number of parameters of the main network, thereby preserving real-time performance. To further enhance visual representation, a dual-scale text generation module is proposed to describe hand features from both global and local perspectives. Specifically, the global text prompt models the overall hand pose based on the bending degree of each finger, while the local text prompt describes local hand features according to the spatial positions of individual joints. In addition, contrastive learning is employed to enforce consistency between multi-scale text features and image features in a shared semantic space. Considering that the CLIP model is highly sensitive to textual formulation, manually designing prompts usually requires extensive tuning and still cannot guarantee sufficient alignment with image features. To this end, this paper adopts a combination of fixed text prompts and learnable word vectors, where the fixed prompts are used to summarize the main semantic information, and the learnable word vectors are used to adaptively refine the prompts, thereby improving the suitability of the text descriptions for the hand mesh reconstruction task. Experimental results show that, compared with real-time methods, the proposed method achieves excellent reconstruction accuracy while maintaining real-time performance. On the FreiHAND dataset, the PA-MPJPE and PA-MPVPE reach 5.5 mm and 5.8 mm, respectively; on the DexYCB dataset, they reach 5.4 mm and 5.2 mm, respectively. The inference speed reaches 68 fps. Ablation studies further demonstrate that the dual-scale text prompts play a key role in hand mesh reconstruction.

With the proliferation of digital platforms, the forms of offensive memes have become increasingly complex and diverse. This phenomenon has exacerbated the scarcity of high-quality annotated data, making modal semantic alignment bias under small-sample conditions a core issue constraining detection performance. To address this issue, this study proposes an offensive meme detection method via Cross-Modal Meta-Learning with Unimodal Rectification（CMML-UR）. This method uses a cross-modal dual-gradient meta-learning framework, leverages hierarchical image features to provide multi-level visual semantics, and combines them with low-noise textual representations generated through multi-regularized modeling. At the decision fusion stage, by evaluating the output confidence of each modality at the sample level, the method introduces a unimodal confidence-gated rectification mechanism to dynamically calibrate the final prediction. Experimental results on the MultiOFF dataset demonstrate that the proposed method achieves a weighted F1-score of 74.6%, which is an improvement of 4.3 percentage points over the state-of-the-art (SOTA) model. In few-shot generalization tests, it maintains a weighted F1-score of 69.3% (5.6 percentage points higher than the baseline model at 63.7%), verifying its efficiency in complex cross-modal semantic understanding and robustness in noise suppression within few-shot scenarios.

This paper addresses the dual challenges of system heterogeneity and data heterogeneity faced by federated learning in the medical image classification task, and proposes an adaptive federated optimization framework named SEFedProX based on reinforcement learning. This framework employs the Soft Actor-Critic algorithm in an heterogeneous environment, based on key state features such as client data distribution and performance feedback, and dynamically adjusts the proximal term coefficients in the continuous action space, effectively overcoming the quantization errors and model oscillation problems caused by discrete action spaces, and achieving precise and smooth control of the local training intensity. At the same time, an EfficientNetV2B2 pre-trained on ImageNet is introduced as the feature extraction network, which improves the model's representation efficiency and discrimination ability while significantly reducing the deployment requirements for resource-constrained medical edge devices, alleviating the overfitting risk in small-scale medical data. Systematic experimental results based on four different system heterogeneity settings and four medical image datasets and a general dataset show that SEFedProX significantly outperforms existing baseline methods in terms of classification accuracy, convergence speed, stability, and robustness. Ablation experiments further verify the effectiveness of the SAC continuous regulation mechanism and the EfficientNetV2B2 network, as well as their collaborative enhancement effect in the framework. This research provides a stable, efficient, and highly adaptive technical solution for the construction of distributed intelligent diagnostic systems in heterogeneous medical environments.

Multimodal vision-language foundation models show great potential in the medical domain, yet face notable limitations due to complex medical semantics and challenging cross-modal modeling. Patient-level rigid alignment ignores semantic similarity, causing unreasonable negative repulsion and degrading learning, while the lack of unified hierarchical modeling between reports and images hinders fine-grained cross-modal alignment. To address the above issues, this paper proposes a global-local collaborative alignment (GLCA), which achieves an improved medical vision-language classification model. GLCA consists of two main components: semantic-driven cross-patient soft global alignment and progressive three-granularity intra-patient local alignment. The semantic-driven cross-patient soft global alignment leverages cross-patient semantic sample pair mining and correlation-weighted contrastive penalty to construct a more continuous feature space that better reflects authentic semantic relationships. The progressive three-granularity intra-patient local alignment aligns visual and textual features at three levels-coarse (report-image), mid (sentence-region), and fine (word-patch)-via progressive query fusion, enabling effective cross-modal interaction. Global-local collaborative alignment first builds a semantically consistent feature space through inter-patient soft global alignment, then performs layer-wise matching via intra-patient multi-granularity alignment, ensuring continuous and precise cross-modal semantic correspondence. Extensive experiments are conducted on four chest X-ray datasets. The results demonstrate that GLCA significantly outperforms existing methods in both zero-shot prediction classification and few-shot fine-tuning classification tasks. On the public 14-class ChestXray14 dataset, the zero-shot prediction classification achieves improvements of 1.2%, 2.0%, and 2.2% over the second-best method in terms of AUC, F1, and ACC, respectively.

Multimodal emotion recognition in conversations integrates language, acoustic, and visual information to automatically identify the emotions in dialogues, thereby enhancing the naturalness and emotional understanding in human-computer interaction. However, existing methods have limitations in modeling multi-layer contextual dependencies of emotions. Multimodal feature fusion often introduces redundant information and noise, and these methods cannot effectively capture the uncertainty of emotions, which limits the recognition of complex emotional categories. To address these issues, this paper proposes a multimodal emotion recognition model that combines hybrid encoding and fuzzy modeling. The model uses a hybrid encoding module to capture both global dialogue context and local utterance-level dependencies, which strengthens the representation of emotional temporal features. In addition, a hierarchical gated fusion mechanism integrates features from different modalities and layers with dynamic weighting to suppress redundancy and noise and improve multimodal feature discrimination. For emotion classification, a fuzzy neural network initialized with linearly spaced parameters models the boundaries of emotion categories using fuzzy membership functions, capturing the uncertainty and fuzziness of emotional expression. Experimental results show that the proposed model outperforms baseline methods on all metrics across the IEMOCAP, MELD, and CMU-MOSEI datasets. It achieves an accuracy of 72.67% on IEMOCAP, 67.37% on MELD, and 54.96% for 7-class accuracy and 86.78% for 2-class accuracy on CMU-MOSEI, respectively, which validates the effectiveness of the proposed method in multimodal sentiment analysis.

Handwritten mathematical expression recognition is an important task in computer vision and plays a significant role in intelligent education, industrial applications, and related fields. Existing encoder-decoder-based methods typically rely on standard convolutions and conventional attention mechanisms for feature extraction. However, the fixed-grid sampling of standard convolution cannot effectively adapt to the geometric deformations of handwritten symbols, which often leads to confusion between visually similar characters. In addition, traditional attention mechanisms usually involve limited cross-dimensional interaction, making it difficult to capture long-range structural dependencies in complex mathematical expressions. To address these issues, this paper proposes a handwritten mathematical expression recognition model based on an encoder-decoder architecture, termed DDTAFF, which integrates deformable dilated convolution and triplet attention feature fusion. Specifically, deformable dilated convolution incorporates learnable dilation rates into both the offset learning process and the customized convolution operation of deformable convolution, enabling more accurate offset prediction and adaptive expansion of the receptive field. Meanwhile, triplet attention feature fusion adopts a similarity-guided dynamic fusion strategy to enhance cross-dimensional feature interaction and improve the extraction of discriminative features. In the encoder, deformable dilated convolution is used to capture multi-scale features and broader contextual information, while triplet attention feature fusion effectively fuses features at different levels to strengthen the representation of critical regions. In the decoder, a Transformer-based structure is introduced to enhance long-range dependency modeling. Experimental results on the CROHME 2014, CROHME 2016, CROHME 2019, and HME100K datasets show that the proposed model achieves recognition accuracies of 59.34%, 59.77%, 59.63%, and 68.94%, respectively, representing improvements of 2.34%, 3.71%, 4.75%, and 1.63% over the baseline model. These results demonstrate the effectiveness and superiority of the proposed method.

In Industrial Internet of Things (IIoT) scenarios, Time-Sensitive Networking (TSN) is required to guarantee high reliability, determinism, and low latency for data transmission. However, the traditional Cyclic Queuing and Forwarding (CQF) model faces challenges such as severe resource preemption, load imbalance, and insufficient network resource utilization when processing mixed traffic. To address these issues, this paper proposes A Deep Noisy Q-Network-Based Multi-CQF Scheduling Algorithm (DNQN-MCQF). Initially, the algorithm constructs a four-channel Multi-Cyclic Queuing and Forwarding (Multi-CQF) architecture, establishing a dedicated queue to guarantee the deterministic transmission of high-priority traffic, and utilizing weights determined by Particle Swarm Optimization (PSO) to calculate dynamic sorting scores for optimizing the transmission sequence. Subsequently, a hybrid feature extraction framework is constructed, employing a Convolutional Neural Network-Gated Recurrent Unit (CNN-GRU) to capture spatiotemporal dynamic features of slot utilization and queue states, combined with a Graph Convolutional Network (GCN) to aggregate static global topological information, generating state embeddings via an Attention Fusion Mechanism. Ultimately, in the scheduling decision phase, the Deep Noisy Q-Network is integrated to inject parameter space noise, achieving adaptive policy exploration and slot optimization. Simulation results show that under different network topologies and time slot conditions, DNQN-MCQF improves the average scheduling success rate and load balance by 14.7% and 19.2%, respectively, compared with baseline algorithms.

Low-light image enhancement (LLIE) is crucial in computer vision for restoring rich visual information from corrupted low-light images. However, existing LLIE methods often suffer from color bias due to color space sensitivity, and they typically fail to balance denoising and color fidelity within a single-stage framework. To address these challenges, this research introduces a novel Dual-Stage HVI-based Transformer Network (DHTNet) for LLIE. DHTNet significantly improves the quality of low-light images by decoupling intensity (I) and color (HV) maps, enabling their independent yet synergistic optimization. In the first stage, a hierarchical Transformer network equipped with an Adaptive Guidance Interaction Module (AGIM) models long-range dependencies between I and HV features. This stage achieves global noise suppression and accurate color calibration. In the second stage, the Multi-Scale Enhanced Synergistic Attention (MESA) module enhances localized color and feature representation through synergetic optimization across I and HV branches. This dual-stage framework addresses the limitations of existing LLIE approaches by retaining complex image details while enhancing visual realism. Experimental results show that DHTNet achieves the highest PSNR on both the SICE and SID datasets, surpassing the second-best method by 0.717 dB and 1.897 dB, respectively. In addition, DHTNet attains PSNR values of 28.756 dB, 24.683 dB, and 25.950 dB on the LOLv1, LOLv2-Real, and LOLv2-Synthetic datasets, respectively, outperforming existing methods such as Retinexformer, CIDNet and other models.

Backdoor attacks on deep neural networks manipulate model behavior by implanting stealthy triggers into the training data, posing a serious threat to model security. However, most existing imperceptible backdoor attacks focus only on invisibility in the spatial domain while overlooking anomalies in frequency-domain features. These methods often introduce noticeable high-frequency artifacts or stable spectral residual patterns in the frequency domain, making them vulnerable to detection by frequency-based defense techniques. To address this issue, a dual-domain imperceptible backdoor attack method based on wavelet packet decomposition and fast Fourier transform is proposed. First, wavelet packet decomposition is employed to select carrier sub-bands according to the energy distribution characteristics of the target class, and an energy-aware adaptive trigger embedding strategy is applied to balance attack effectiveness and stealthiness. Then, fast Fourier transform is used for spectral reconstruction by combining the amplitude spectrum of clean samples with the phase spectrum of poisoned samples, thereby reducing detectable traces in the frequency domain. Comparative experiments are conducted on the CIFAR-10, CIFAR-100, and Tiny ImageNet datasets using PreAct-ResNet18 and VGG19-BN models. The results show that the proposed method maintains high attack effectiveness while significantly improving dual-domain stealthiness and robustness against defenses. On CIFAR-10, it achieves an attack success rate of 99.94% and demonstrates strong evasive capability against tested defenses such as FTD, Neural Cleanse, and STRIP.

Liver segmentation is an important prerequisite for liver disease diagnosis, three-dimensional reconstruction, and surgical planning. To overcome the segmentation challenges caused by complex structure, weak boundary, and large individual differences of liver in abdominal CT volumes, an automatic segmentation method based on multi-view information fusion is proposed. Firstly, a 2D U-shaped convolution network based on atrous spatial pyramid pooling is designed to extract multi-scale features and enlarge the receptive field without increasing the number of network parameters. The proposed 2D U-Net is then applied to slice-wise segmentation from multiple view directions, including axial, sagittal, and coronal planes, thereby compensating for the inability of single-view models to capture inter-slice contextual information. Subsequently, a lightweight 3D convolution network is built to fuse the segmentation results on multiple viewing directions, enabling three-dimensional liver segmentation under limited computational resources and yielding voxel-wise liver probability maps and label assignments for the CT volumes. Finally, the probabilities and labels obtained are used to construct the graph cut energy function to refine the segmentation results, effectively alleviating over-segmentation and under-segmentation. The proposed method can extract 3D features of CT volume indirectly by fusing the segmentation results obtained on different viewing directions and can improve the segmentation accuracy by introducing graph cuts. Experiments are carried out on two public datasets including 3DIRCADb and LiTS, and the Dice obtained by the proposed method on the test set are 0.947 and 0.962 respectively, outperforming those by many existing segmentation methods.

To address the issues of unbalanced multi-scale feature expression, cross-level fusion loss, and insufficient bounding box localization accuracy in document detection, a document detection method with multi-scale feature and semantic optimization is proposed. This method includes three parts of design and improvement: first, a multi-branch convolutional attention fusion module is constructed, which expands the receptive field via multi-scale strip convolution and integrates the attention mechanism with the C3k module; second, a multi-scale neck coordinated with global semantics and high-order correlation is designed, which achieves fusion through global feature collection, hypergraph convolution-based correlation mining, and multi-scale scattering; third, the bounding box regression loss is optimized, and dual-threshold interval mapping is adopted to enhance the discrimination of sample losses. Experimental results on the EXAM, CDLA, D4LA, and PubLayNet datasets show that the average detection accuracy of this method is significantly higher than that of existing methods. Experimental results indicate that this method can break through the performance bottleneck of YOLO11n in the field of document detection, improve accuracy while ensuring efficiency, and provide a scientific and feasible application scheme for document detection.

To address the challenges of low-rank degradation and partial observation in spatio-temporal data caused by unstructured occlusion, motion distortion, and multi-source noise coupling in complex dynamic scenes, this paper proposes a motion discontinuous spatio-temporal behavior understanding framework that integrates non-convex temporal difference low-rank constraints and hierarchical trajectory-behavior semantic mapping. Firstly, a temporal difference low-rank recovery model based on the non-convex Schatten-p norm is constructed, and the Alternating Direction Method of Multipliers (ADMM) is employed to reconstruct motion data under high missing rates and noise pollution. Secondly, based on the recovered data, structured trajectory clusters are built by combining multi-object tracking, and trajectory neighborhood interaction features are extracted. Furthermore, a three-level behavior understanding model is proposed: behavior primitive classification based on multilayer perceptrons, interaction pattern recognition based on graph attention networks, and semantic fusion and behavior narrative generation incorporating spatio-temporal context, achieving end-to-end mapping from trajectories to high-level semantics. Experiments show that the proposed method significantly outperforms baseline approaches in data recovery quality under a 60% high missing rate, achieving behavior recognition accuracies of 92.7% on both the NTU RGB+D (X-Sub) and the self-built motion dataset BAS, which is 5.6 percentage points higher than the best comparative method. Ablation studies further validate the effectiveness of each module: the NTDLR recovery module improves the recognition rate from 78.3% to 86.7% under 60% missing data, trajectory neighborhood encoding enhances it to 88.2%, and the complete three-level model achieves optimal performance through synergistic interaction. The results of interaction pattern recognition and semantic description generation also notably surpass those of mainstream graph convolutional networks and their variants. This research provides an interpretable and scalable algorithmic framework for discontinuous and interactive motion behavior understanding in complex dynamic scenes.

To address the challenges of low contrast, large-scale variation, and limited computational resources in industrial steel surface defect inspection, this paper proposes an efficient detection network with full-path collaborative enhancement, termed ISA-DETR. First, an Information-Preserving Downsampling (IPD) structure is constructed, which adopts a spatial–channel rearrangement-based pixel reorganization strategy to replace conventional strided downsampling. This design effectively preserves fine-grained spatial information while reducing feature map resolution, thereby alleviating the information loss of small defects during feature extraction. Second, a Large Separable Kernel Attention-Hybrid Group Block (SLK-HG) module is developed by integrating large-kernel separable attention mechanisms. Through the collaborative optimization of group convolution and separable convolution, the module builds an ultra-large receptive field with near-linear computational complexity, significantly enhancing the network’s ability to model long-range spatial dependencies and irregular defect patterns. Furthermore, an Adaptive Dynamic Sampling (ADS) operator is introduced to achieve precise cross-scale feature alignment via content-driven offset prediction, reducing localization deviations in complex backgrounds and improving detection robustness. Experimental results on the NEU-DET steel surface defect dataset demonstrate that ISA-DETR achieves an mAP@0.5 of 75.2% with only 20.67M parameters and 77.5 GFLOPs. Compared with the baseline model, the proposed method reduces the number of parameters and computational cost by 35.4% and 25.1%, respectively, while improving detection accuracy by 3.2%. In addition, transfer experiments on the PCB defect dataset further verify the strong generalization capability of the proposed approach. Overall, the proposed method achieves an effective balance between detection performance and deployment efficiency, providing a practical and reliable solution for intelligent quality inspection in industrial edge scenarios.

Abstract In the high-risk petrochemical industry, operational environments involve complex factors such as high temperature, high pressure, flammable, explosive, and toxic substances. Even minor deviations in worker behavior can lead to severe accidents, resulting in irreversible casualties and property losses. Traditional supervision methods, which rely heavily on manual inspection, are not only inefficient but also struggle to cover multi-worker and multi-equipment collaborative scenarios, and are highly susceptible to subjective interference. The core challenges in recognizing group collaborative behaviors lie in modeling complex human–object interactions, capturing dynamic features of multiple targets, and bridging the ambiguous mapping between macroscopic group intentions and microscopic individual actions. To address these issues, this paper proposes a graph neural network-based method for group collaborative behavior recognition. By constructing a unified interactive graph structure, entities such as workers and equipment are encoded as nodes, and multimodal perceptual features are integrated to enable end-to-end reasoning of interpersonal and human–object interactions. A hierarchical graph network architecture is further designed to model the correlation and evolution from individual actions to group behaviors, achieving accurate recognition and understanding of multi-target group behaviors in complex operational scenarios. Comparative experimental results show that the proposed method improves the MCA/MPCA metrics by 3.91% and 2.86%, respectively, over the next best method on a self-built dataset. On the public open-source Volleyball dataset, the MCA/MPCA metrics are improved by 0.26% and 0.21% compared to the next best method, fully verifying the method's advancement and robustness.

Multi-interest sequential recommendation models extract multiple user interests through a dynamic routing mechanism to achieve personalized recommendations. However, during the interest extraction phase, improper item–interest routing weight allocation may cause multiple interest representations to become excessively similar, leading to the multi-interest collapse issue. In the prediction phase, neglecting users’ preference intensity across different interests may assign comparable or even greater influence to low-preference interests than to high-preference ones, resulting in the multi-interest preference weight imbalance issue. To address these issues, we propose a multi-interest sequential recommendation model based on disentangled feature representation and adaptive weight fusion (DMIAFRec). First, the model partitions items according to their co-occurrence relationships, grouping frequently co-occurring items with complementary semantic characteristics into the same interest group. This partition serves as a structural guidance mechanism for routing weight allocation, encouraging each group to focus on relatively independent user interests, thereby achieving disentangled multi-interest representations and preventing excessive convergence among interest vectors. Furthermore, a time-decay mechanism and a multi-interest attention fusion strategy are introduced to adaptively assign weights according to users’ preference intensity for each interest. The weighted aggregation of multiple interest representations produces a unified user preference representation that reflects heterogeneous interest importance, thus enhancing personalized recommendation performance. Experimental results on three public datasets show that, compared with the best baseline model, the proposed model achieves average improvements of 6.2%, 4.98%, and 4.07% in R@20, NDCG@20, and HR@20, respectively, on the Retail Rocket, Gowalla, and Books datasets, demonstrating its effectiveness in improving recommendation performance and addressing the aforementioned issues.

Deep neural networks have been widely applied to time-series critical tasks such as medical diagnosis, intelligent sensing, and autonomous driving, yet their security vulnerabilities have gradually emerged. Existing studies show that deep time-series models are also susceptible to adversarial attacks. However, most existing adversarial attack methods for time-series models mainly focus on norm-bounded numerical perturbations and often ignore the inherent causal dependencies and dynamic evolution in the data generation process. As a result, the generated adversarial examples may deviate from feasible system dynamics and lack practicality in real-world scenarios. Therefore, generating effective adversarial examples while adhering to temporal causal dynamics has become an important challenge in time-series adversarial research. To address this challenge, this paper proposes TCADE (Temporal Causal ADversarial Examples), a novel method that explicitly models causal structures in time-series data and performs counterfactual reasoning under causal intervention constraints. By formulating adversarial attacks as feasible interventions on the underlying system, TCADE generates adversarial examples that can effectively mislead model predictions while remaining consistent with the system’s causal relationships and dynamic evolution. Experimental results demonstrate that TCADE achieves significant attack effectiveness in black-box settings, and the generated adversarial sequences conform to the causal generation mechanisms. This work provides a systematic evaluation of the vulnerability of time-series models under realistic and feasible black-box attacks, and offers practical insights for improving model robustness.

Federated learning is a distributed machine learning paradigm that leverages decentralized data resources while ensuring data privacy. However, in real-world scenarios, data across clients are often non-IID (Independent and Identically Distributed), leading to label shift and class imbalance issues, which hinder convergence of global models and degrade generalization performance. To address the impact of such data heterogeneity on model performance, we propose a cross-client data augmentation and classification framework based on diffusion models. In this framework, each client trains an initial diffusion model based on local data and uploads its model parameters to the server. The server aggregates these parameters to construct a global diffusion model, which is then downlinked to all clients. Clients use the global diffusion model to generate supplementary samples, which are uploaded to the server for data augmentation to balance the local class distribution, thereby improving classifier performance. Ultimately, the classification model is trained through federated learning by receiving both local data and generated samples, and is deployed to clients for image classification and recognition. To generate high-quality images, a denoising diffusion probabilistic model is used as the generation backbone, while a ResNet-18 architecture is employed for the federated classification model. Experimental results show that the fine-tuned global diffusion model can generate images that are more consistent with the real data distribution. By augmenting the data through generated samples, the local data distribution on clients becomes more balanced, significantly improving global classification accuracy. Under the non-IID condition with a Dirichlet coefficient α=0.1, the accuracy of CIFAR-10 and CIFAR-100 increased from 46.76% and 21.31% to 54.64% and 25.57%, respectively, demonstrating the effectiveness of the proposed data augmentation strategy in mitigating class imbalance.

This paper proposes a 3D digital rock reconstruction framework based on diffusion prior guidance and multi-scale residual fusion Implicit Neural Representation (INR) to address challenges such as structural discontinuity, topological fracture, and the difficulty in balancing cross-scale microscopic details under sparse 2D slice conditions. The framework introduces the Score Distillation Sampling (SDS) mechanism to transform the geometric and topological priors in pre-trained diffusion models into continuous gradient guidance. It combines the measurement consistency loss for collaborative constraints to achieve the consistent restoration of local fine features and global topological structures under extremely sparse slice constraints. Meanwhile, the framework utilizes multi-scale residual structures to enhance the representation ability of INR for complex pores and improves the generalization performance of the model under different voxel sizes. Experimental results show that this method accurately restores complex pore spaces on various digital rock datasets. The reconstructed structures maintain high consistency with the ground truth in key physical indicators such as porosity distribution and geometric connectivity. In the 256⊃3; scale reconstruction task, the Dice Similarity Coefficient (D_dice) reaches 97.01%, which is 2.6% higher than the baseline INR model. As the reconstruction scale further expands, the D_dice maintains at 95.97% and 92.88% under 512⊃3; and 1024⊃3; high voxel size tasks, respectively, demonstrating excellent stability in large-scale reconstruction. In the cross-sample generalization tests for Berea sandstone and Ketton limestone, the D_dice reaches 93.44% and 95.51%, respectively. This study solves the problem of stable reconstruction for complex porous media in data-limited scenarios and provides a physically reliable and continuous new technical solution for refined geological modeling.

Accurate short-term power load forecasting is crucial for the safe operation and optimal scheduling of power systems. Existing forecasting methods based on decomposition rely on fixed prior knowledge, which leads to rigid decomposition patterns. As a result, they struggle to handle load data with multiple periodicities and strong non-stationarity. Meanwhile, these methods find it difficult to balance computational complexity and prediction accuracy. To address these issues, this paper proposes a forecasting model based on learnable wavelet decomposition and KAN-Mixer, called LWKAN-Mixer. First, a learnable wavelet decomposition module is used to decompose the original load sequence into wavelet components of different frequency bands. Next, the Fast Fourier Transform (FFT) is applied to extract the dominant period of each component. Based on these dominant periods, patches of corresponding sizes are created for each component. Then, a multi-scale time-frequency fusion module is used to model each component independently to capture time-frequency features. The KAN-Mixer and a dual interactive convolution block are used to capture sequence representations and temporal dependencies, respectively. A multi-scale hybrid loss function is also introduced to constrain the quality of decomposition and reconstruction during training. This helps alleviate error accumulation and improve prediction accuracy. Experimental results on three real-world load datasets show that the proposed model reduces MAE by 1.10%–9.37% on the Australia dataset and by 4.97%–17.36% on the Morocco dataset compared to the latest baseline models. On the Cele dataset, the model achieves the second-best MAE. These results demonstrate that LWKAN-Mixer can effectively model complex nonlinearity and non-stationarity in load sequences, achieving strong performance in short-term load forecasting tasks.

With the continuous advancement of social media and online service platforms, user reviews have increasingly become a critical source of information influencing consumer decisions and product evaluations. Aspect-based sentiment analysis (ABSA), as a key research direction in fine-grained sentiment computing, still faces significant challenges in practical applications—particularly prominent semantic ambiguity in textual content and insufficient extraction of sentiment cues. To address these limitations, this paper proposes CKMA, a novel aspect-based sentiment analysis model that integrates knowledge graph embeddings with a multi-channel attention mechanism.The CKMA model first leverages knowledge graph embedding techniques to map entities and their relationships from external knowledge bases into low-dimensional semantic vectors, which are then fused with textual representations to alleviate semantic ambiguity commonly observed in user reviews. Building upon this knowledge-enhanced representation, we design a parallel multi-channel feature extraction framework comprising three distinct channels: a structured information channel, a context-aware channel, and an aspect-focused channel. Through a staged fusion strategy, this framework enables collaborative modeling of diverse semantic and syntactic signals, thereby enhancing the model’s capacity to capture aspect-relevant sentiment features. To mitigate the loss of original semantic information during deep feature learning, we further introduce a joint fusion mechanism that combines the knowledge-enhanced word-level representations with the outputs of the multi-channel attention modules, thereby improving the completeness and robustness of the final feature representation.Extensive experiments conducted on four widely used benchmark datasets—Restaurant14, Restaurant16, Laptop14, and Twitter—demonstrate that the proposed method achieves superior performance in terms of both accuracy and Macro-F1 scores. Notably, CKMA exhibits more pronounced advantages on datasets characterized by complex syntactic structures, validating the effectiveness of our synergistic modeling strategy that jointly exploits structural and semantic information for aspect-based sentiment analysis.