15 July 2026, Volume 52 Issue 7
-
Select all|
-
Computer Engineering. 2026, 52(7): 1-21. https://doi.org/10.19678/j.issn.1000-3428.0260431
A considerable structural contradiction exists between the surge in the parameter scale of Large Language Models (LLMs) and the limited physical resources of edge terminals, which restricts their large-scale deployment. Traditional cloud-centric inference relies heavily on network transmission and suffers from high communication latency, making it difficult to meet the dual demands of ultra-low latency and strict data privacy protection in scenarios such as autonomous driving and intelligent healthcare. Meanwhile, edge physical hardware, ranging from microcontrollers to edge gateways, exhibits significant heterogeneity, making it difficult to transfer of general cloud-side compression schemes directly. To this end, grounded in the heterogeneous physical constraints of edge devices, this study systematically reviews the technical system of efficient compression and software—hardware co-deployment of large models on the edge. First, this study analyzes the underlying mechanisms of three core compression techniques, namely model quantization, parameter pruning, and Knowledge Distillation (KD), specifically in edge scenarios. Although Post-Training Quantization (PTQ) offers deployment agility, it faces the challenge of representation collapse caused by abnormal long-tail activations in large language models. While more robust, Quantization-Aware Training (QAT) is constrained by the scarcity of retrained computations at the edge. This study demonstrates the practical energy efficiency advantages of structured pruning on hardware with constrained memory bandwidth, highlighting that the high theoretical compression ratio of unstructured pruning is easily offset by the indexing and addressing overhead on general-purpose edge chips. Traditional shallow parameter alignment risks feature loss and bias amplification when bridging the capacity gap between teacher and edge student models. In summary, the single-compression techniques exhibit clear diminishing returns under extreme constraints. Second, to alleviate the performance bottlenecks of individual techniques, this study generalizes a multi-stage hybrid compression paradigm driven jointly by model architecture and physical scenarios. The three core optimization pipelines are systematically categorized as follows: 1) a serial pipeline strategy that pursues a high physical compression rate and suits real-time inference on edge gateways; 2) a deeply coupled joint optimization flow that strictly balances energy efficiency and accuracy by integrating quantization, pruning, and low-rank decomposition within a single framework for synchronous updates, which is suitable for power-constrained mobile terminals; and 3) a distillation-driven mechanism for deploying large-parameter models that uses teacher priors to guide structural reshaping and quantization. This multi-stage paradigm effectively expands the multi-dimensional trade-off space among model scale, computational power consumption, and fidelity. Furthermore, facing a broad spectrum of computing power and energy consumption levels, this study constructs a four-layer software—hardware co-design mechanism comprising the system, model, operator, and instruction levels. It clearly states that the focus of co-optimization must shift dynamically according to physical substrate constraints. At the system level, it focuses on resource-aware scheduling and task distribution in cloud—edge environments; at the model level, it relies on hardware-aware neural architecture search to achieve structural adaptation; at the operator level, it promotes cross-layer fusion and memory access locality restructuring; and at the instruction level, it targets custom extended instructions for specific microarchitectures (e.g., RISC-V) to precisely control underlying energy consumption. Together with the full-chain deployment process comprising model conversion, compilation reengineering, and memory management (e.g., SwapNet), this mechanism effectively maps compression algorithms to the underlying physical execution and improves the comprehensive utilization efficiency of heterogeneous computing resources. Finally, future research challenges in the field of edge intelligence lightweighting are prospectively identified. It emphasizes that robust compensation mechanisms for ultra-low bitwidth (≤4 bit), hardware-adaptive dynamic semi-structured pruning, and effective knowledge transfer for deep logical reasoning in large models are core directions for overcoming current lightweighting bottlenecks. Simultaneously, a hardware-agnostic unified toolchain based on deep-learning compilers must be built to eliminate deployment barriers caused by fragmented heterogeneous devices. This systematic technical review provides solid theoretical support and a reference guide for developing a low-latency, strong-privacy protection edge intelligence ecosystem.
-
Computer Engineering. 2026, 52(7): 22-39. https://doi.org/10.19678/j.issn.1000-3428.0070189
The rapid development of blockchain technology and the rise of cryptocurrencies not only challenge traditional financial and asset management concepts but also introduce significant privacy and security concerns. Coin mixing, a primary privacy protection mechanism in blockchains, is designed to enhance transaction anonymity and protect user identities. However, this enhanced anonymity also enables the misuse of cryptocurrencies for illicit activities such as money laundering, theft, and fraud. Therefore, a comprehensive review of technologies for the identification, detection, and regulation of coin mixing services is urgently required to support law enforcement and promote the secure and compliant use of cryptocurrencies. This paper first reviews the technical principles and mechanisms of coin mixing services and proposes a general classification framework for mixing tools while discussing the criminal activities associated with them. Subsequently, it focuses on recent advances in five key regulatory approaches: manual rule-based methods, address clustering, entity classification, anomalous transaction detection, and mixing service traceability. It also surveys existing industrial tools and platforms related to coin mixing regulations. Finally, it identifies current challenges in the field and suggests directions for future research.
-
Computer Engineering. 2026, 52(7): 40-55. https://doi.org/10.19678/j.issn.1000-3428.0070700
Deep learning has achieved significant success in computer vision and speech signal processing. However, its rapid development of deep learning has had negative effects. Many types of fake videos and voices are flooding the Internet. Some criminals use deep learning technology to replace the face in original video, edit facial attributes, and synthesize or clone the speaker's voice. Criminals can cause social unrest and chaos by producing pornographic videos, fake news, and political rumors, which threaten personal interests and national security. Many scholars have proposed solutions to eliminate these negative effects. Early forgeries focused primarily on single-modal forgeries. Therefore, most current solutions focus on single-modal counterfeit detection and fail to fully consider the intrinsic relationship between audio and video. Existing single-modal detection methods often exhibit suboptimal performance when both audio and video are forged. Recently, with the increase in research, some scholars have begun to explore the use of multimodal models for counterfeit detection and have achieved remarkable results. This survey reviews video and voice forgery and detection technologies, collects and sorts voice, audio, and video forgery datasets, and summarizes multimodal counterfeit detection methods. Finally, the existing problems and future research directions of the current detection technology are analyzed.
-
Computer Engineering. 2026, 52(7): 56-75. https://doi.org/10.19678/j.issn.1000-3428.0070621
Cyber Threat Intelligence (CTI) provides critical support for the construction of cybersecurity defense systems by analyzing and predicting cyber threats. Information Extraction (IE) technology, which is a foundational component of CTI, has garnered significant attention from both academia and industry in recent years. Its application is particularly important for constructing threat intelligence knowledge graphs. Given the ever-evolving nature of threat types and attack techniques, precise IE capabilities play a decisive role in enhancing cybersecurity defense. This review systematically examines the development of IE technologies within the CTI domain and presents a comprehensive analysis and summary of the key components of IE frameworks. IE methods are categorized into three major types: traditional machine-learning-based, deep-learning-based, and large language-model-based methods. The characteristics, applicable scenarios, and progress of research in each category are discussed in detail. Moreover, quantitative and theoretical analyses of existing techniques are presented, along with a discussion of their implications and limitations in practical applications. The review also outlines the current challenges in the field and proposes potential research directions to address them. By presenting a detailed assessment of IE technologies for CTI, this review offers a comprehensive academic perspective and lays an important theoretical foundation for fostering more efficient and intelligent cybersecurity defense systems.
-
Computer Engineering. 2026, 52(7): 76-85. https://doi.org/10.19678/j.issn.1000-3428.0070737
As a key technology in multi-agent game confrontation, opponent modeling aims to learn the behavior of the opponent to reduce the uncertainty of the environment and facilitate decision-making. However, most existing opponent modeling methods adopt the structure of offline training and online adaptation. During offline training, traditional neural dynamics model is used to predict the agent step-by-step, which facilitates the formation of single-step and cumulative errors. Additionally, when facing an unknown opponent during online adaptation, the planned states of the controlled agent deviate from the distribution of the dataset. To solve these problems, a method based on a diffusion model and cross-attention is proposed to establish a correlation with the opponent. The cumulative bias problem is solved using the feature that the diffusion model can generate multistep planning sequences simultaneously. The concept of a strategy set is proposed, and the deviation problem is not only solved by online fine-tuning but also prevents the problem that the offline strategy will be destroyed in the initial stages of online training. Experimental results in both open intensive reward and sparse reward competitive environments demonstrate the superior performance of this method.
-
Computer Engineering. 2026, 52(7): 86-93. https://doi.org/10.19678/j.issn.1000-3428.0070459
Several recent studies have focused on the field of Class Incremental Learning (CIL). A common strategy involves freezing the parameters of the feature extractor after the initial training phase of the model for maintaining the stability of the learned knowledge. This strategy relies on building a robust feature extractor during the initial training phase. However, this method may not reach its full potential in scenarios with limited data during the initial phase, resulting in poor incremental learning. To optimize the initial training in incremental learning and cope with the challenge of limited data at the initial stage, a two-step training method that is the initial phase training method of CIL based on class boundary sample generation and fine-tuning, is proposed for the initial stage of feature extractor training. The method comprises two steps: generative adversarial training and fine-tuning training. Generative adversarial training uses a generative adversarial network to generate samples near the boundary of each class and uses contrastive learning to force the center of the class away from these generated samples for enhancing the discrimination of features between different classes and improving the generalization performance of the model. In the fine-tuning stage, cross-entropy loss and self-supervised loss are combined to fine-tune the model and improve the classification accuracy on the initial task. The resulting trained feature extractor is used as the starting point for other incremental learning methods. The method improves the average accuracy of FeCAM by 3.8 percentage points and FeTrIL by 2.2 percentage points on the CIFAR-100 dataset with 50 base classes and five incremental tasks.
-
Computer Engineering. 2026, 52(7): 94-106. https://doi.org/10.19678/j.issn.1000-3428.0070584
In recommender systems, sequential recommendations aim to predict the future interests of a user based on historical interaction sequences. However, existing deep learning models typically focus on capturing users' long-term behavior patterns while neglecting the fine-grained modeling of temporal information, which limits improvements in recommendation performance. To address this issue, this study proposes a method that combines the global and local self-attention mechanisms. This method processes user interaction sequences in chunks and introduces a weight decay mechanism that assigns differentiated weights to different sequence blocks based on temporal distance, thereby capturing changes in users' short- and long-term interests more accurately. However, this approach introduces additional parameters, thereby increasing the model complexity. While Stochastic Shared Embedding (SSE) technology can reduce the parameter count and mitigate overfitting, its random embedding approach may introduce noisy data, affecting the accuracy of the recommendations. To solve this problem, this study proposes a strategy that combines a Generative Adversarial Network (GAN) with SSE. By using GAN to generate high-quality interaction data that align with user interest distributions and combining this with the random replacement mechanism in SSE for data augmentation, the data are enhanced by randomly selecting generated data, thus reducing the risk of noise introduction while retaining SSE's advantage in reducing overfitting. Experiments are conducted on three public datasets: MovieLens-1M, Amazon Beauty, and Yahoo Music. The results show that the proposed method performs well in terms of Normalized Discounted Cumulative Gain (NDCG), Hit Rate (HR), and Mean Reciprocal Rank (MRR).
-
Computer Engineering. 2026, 52(7): 107-120. https://doi.org/10.19678/j.issn.1000-3428.0070533
Semi-supervised text classification uses a limited set of labeled data in conjunction with a large corpus of unlabeled data to develop text classification models. Current pseudo-labeling techniques often face challenges such as pseudo-label bias stemming from decision boundary underfitting and error accumulation. To address these issues, a cross-annotation model based on the multi-granularity semantic margin loss is introduced. Initially, the Bidirectional Encoder Representations from Transformers (BERT)-pretrained model is employed to extract vector sequences enriched with contextual information. Subsequently, an attention mechanism and Text Convolutional Neural Network (TextCNN) are utilized to capture global and local semantic features, respectively. These features are subsequently fused to create enriched semantic representations, factoring in semantic similarities across different categories to separate similar classes within the feature-embedding space, thereby reducing the impact of error accumulation. Additionally, a category-specific margin to calculate the positive and negative sample loss between pseudo-labels derived through adaptive local threshold learning is introduced. The margin enables a high-density distribution of samples within each class in the embedding space, thereby mitigating decision boundary underfitting. Experimental evaluations demonstrate that the proposed model effectively integrates multi-granular semantic information, achieves high cohesion among pseudo-labeled samples, and significantly improves the performance of semi-supervised text classification.
-
Computer Engineering. 2026, 52(7): 121-130. https://doi.org/10.19678/j.issn.1000-3428.0070456
Event extraction is a crucial task in the field of information extraction that aims to identify and extract specific events or factual information from natural language texts. A key challenge in event extraction is event overlap, where a single word may act as a trigger for different event types or play varying roles across events. Existing methods for overlapping event extraction often fail to adequately consider the relationships between argument roles, which affect the event extraction performance. To address the nested event challenge in the event extraction domain, this study proposes a cascade type event extraction model called Argument Association Graph (AAG)-based Cascade Type Event Extraction model (AAGCTEE). AAGCTEE utilizes XLM-RoBERTa-base for deep encoding and enhances the representation between word pairs using conditional layer normalization and dilated convolution techniques. In the trigger recognition module, the AAGCTEE Cascade Type Predictor (CTP) accurately identifies triggers related to multiple event types, effectively solving the common issue of nested triggers in traditional event extraction. In the argument recognition and classification module, AAGCTEE employs an Argument Association Graph (AAG) and a Global Normalization Decoding (GND) strategy to efficiently handle complex nested argument structures. The average F1 values for Trigger Identification (TI), Trigger Classification (TC), Argument Identification (AI), and Argument Classification (AC) of AAGCTEE are higher than those of the comparative models by 4.12, 0.17, 13.66, and 0.98 percentage points for Chinese datasets DuEE and FewFC and English datasets PHEE and CASIE, respectively. Compared with the ablation experiments lacking CTP, AAG, and GND, AAGCTEE demonstrates an average increase of 5.56, 5.31, and 5.08 percentage points in the average F1 values across the four evaluation metrics, verifying its effectiveness in extracting complex events.
-
Computer Engineering. 2026, 52(7): 131-142. https://doi.org/10.19678/j.issn.1000-3428.0070335
Time series anomaly detection is a critical task in the finance, medical treatment, and industrial monitoring fields. However, when handling time series data, traditional methods often face challenges such as limited feature extraction ability, poor generalization ability, and poor real-time performance. These methods may even ignore the temporal relationships between data. To fully consider the temporal dependencies in time series data and further enhance the ability to detect anomalies, this paper proposes an unsupervised time series anomaly detection method called β-VAE-BiLSTM. First, the Variational Autoencoder (VAE) network structure is customized for data reconstruction in anomaly detection. Next, a hyperparameter, β, is introduced to control the weight of the Kullback-Leibler (KL) divergence term in the evidence lower bound, enhancing the encoder module's disentanglement and expressive capabilities in the latent space and obtaining more robust data representations. Then, Bidirectional Long Short-Term Memory (BiLSTM) is used to estimate the long-term correlations of the latent representations, capturing their forward and backward dependencies and performing temporal predictions. Finally, the reconstruction error is calculated by averaging the fusion of the prediction results and the decoder reconstruction. The optimal threshold for anomaly detection is determined through grid search. Experimental results show that the proposed method has superior anomaly detection performance on multiple public time series datasets. It can effectively extract complex temporal data features, achieve efficient computation and real-time anomaly detection, identify anomalous points and regions, improve detection accuracy and robustness, and reduce false positives and false negatives.
-
Computer Engineering. 2026, 52(7): 143-154. https://doi.org/10.19678/j.issn.1000-3428.0070453
Multi-behavioral recommendations have become a research hotspot for personalized recommendation tasks. To address the issues of biased information transfer between front and back behaviors and low differentiation between samples in existing multi-behavioral recommendation models, a Cross-Behavior Feature Regularization and Interaction Contrastive Learning based multi-behavior recommendation model (CBFRICL) is proposed. Cross-behavioral feature regularization is used to improve the inter-behavioral feature transfer mechanism based on the differences in pre- and post-behavioral features. Redundant information from previous behaviors is filtered to reduce the enforced dependence between consecutive behaviors. Therefore, the diversity and relative independence of information across different behaviors are preserved. The learning process of the model is constrained using regularization terms. Feature interference is prevented during the transfer process, reducing information distortion and mitigating error accumulation in cascaded Graph Convolutional Network (GCN) structures. An interaction contrastive learning method is proposed based on the spatial distribution rules of user-item interactions. Positive and negative sample pairs are generated using information on user interaction history. Positive samples are brought closer together, and negative samples are pushed away in the feature space, aiding the model in defining the sample boundaries. The sample clustering effect in the feature space is enhanced by capturing the personalized preferences of each user. Thus, the accuracy of the item recommendations is improved. The experimental results show that on the public datasets Tmall, Beibei, and Jdata, CBFRICL improves the HR@K and NDCG@K metrics by a minimum of 4.88% and 1.58%, respectively, compared to the CRGCN model.
-
Computer Engineering. 2026, 52(7): 155-167. https://doi.org/10.19678/j.issn.1000-3428.0070620
In order to solve the problem of data sparsity in crowdsourcing task recommendation and improve the accuracy of crowdsourcing task recommendation, this paper proposes a crowdsourcing task recommendation method based on collaborative knowledge graph and hybrid neural network. This method first utilizes task entity alignment to fuse the worker—task bipartite graph and crowdsourcing task knowledge graph to form a worker task collaborative knowledge graph to alleviate the problem of data sparsity; Secondly, a Bidirectional Gated Recurrent Unit (BiGRU) is used to encode multiple paths between workers and tasks. Considering the correlation between paths, attention mechanism is utilized to weight and aggregate the information obtained from encoding multiple paths, in order to more accurately learn workers' preferences and recommend crowdsourcing tasks more accurately; At the same time, graph convolutional networks are used to capture the high correlation between crowdsourcing tasks and fully consider the complex semantic information of entities; Finally, based on the embedded representations of the workers and tasks obtained, provide comprehensive recommendations to the workers. The experimental results on six publicly available datasets, MovieLens-1M, Yelp, Book-Crossing, Music, Zhu-Bajie, and CHI, showed that compared with the benchmark model, the method proposed in this paper improved the Aera Under the ROC Curve (AUC) index by an average of 5.8%, 7.85%, 5.75%, 6.3%, 5.47%, and 4.58%, and the model also improved in other indicators. The experimental results demonstrate the effectiveness and stability of the proposed method, which can provide a research approach for the field of crowdsourcing task recommendation.
-
Computer Engineering. 2026, 52(7): 168-178. https://doi.org/10.19678/j.issn.1000-3428.0070180
Accurate traffic flow prediction is critical for urban transportation; however, abnormal events in traffic flow, such as severe weather or emergency road conditions, present significant challenges in maintaining the accuracy of current traffic flow prediction models. Most existing models fail to accurately capture complex dynamic patterns of traffic data and often require substantial computational resources. To address these issues, this paper proposes a Spatio-Temporal Generative Adversarial Clustering Graph Convolutional Network (CG-STGAN) for abnormal traffic flow prediction. The model uses a generator to capture and simulate normal traffic flow data, while a discriminator compares the generated data with the original data to detect abnormalities. Adversarial training helps improve the ability of the model to identify abnormal traffic flows. In addition, a Cluster Graph Convolutional Network (Cluster-GCN) is introduced to divide the traffic graph into multiple subgraphs, thereby limiting graph convolution neighborhood expansion, reducing memory usage, and improving prediction accuracy. A graph convolutional gated recurrent unit is employed to extract short-term spatiotemporal features, while a Long Short-Term Memory (LSTM) module learns long-term dependencies. The results of experiments demonstrate that CG-STGAN outperforms baseline models in abnormal traffic flow prediction.
-
Computer Engineering. 2026, 52(7): 179-188. https://doi.org/10.19678/j.issn.1000-3428.0070260
Current deep learning methods for precipitation spatial interpolation rely primarily on the spatial relationships between stations and often overlook the effective representation of regional features and their complex dependencies. In particular, this approach does not consider the relationship between precipitation and elevation. To address this issue, GeoPriFormer is proposed as a spatial interpolation method for precipitation that incorporates geographic priors. This method uses a Transformer as the backbone network and enhances feature representation by integrating geographic prior data. This integration improves the accuracy and reliability of precipitation spatial interpolation. An elevation-based multi-head attention submodule is introduced to better understand and adapt to the effects of complex terrain. Additionally, a Laplacian method learns the context-aware vector encoding of geographic coordinates, enhancing the model's understanding of spatial relationships and regional characteristics. Consequently, the feature representation is embedded with geographic priors. Comparative experiments conducted on the Henan, Beijing-Tianjin-Hebei, and Baden-Württemberg (BW) datasets show that GeoPriFormer outperforms the baseline models. For example, in the Beijing-Tianjin-Hebei dataset, the Root Mean Squared Error (RMSE) is reduced by 1.6%, the Mean Absolute Error (MAE) decreases by 5.1%, and the Nash-Sutcliffe's Efficiency (NSE) increases by 1.56%, indicating an improvement over the current dominant approaches.
-
Computer Engineering. 2026, 52(7): 189-198. https://doi.org/10.19678/j.issn.1000-3428.0070722
In the field of clinical medicine, the pathological analysis of patient tissue sections is considered the gold standard for assessing complex diseases. Traditional Super-Resolution (SR) methods often fail to effectively capture fine structures and textures in pathological images, leading to suboptimal reconstruction performance. To address this issue, this study proposes a novel super-resolution generative adversarial network based on a parallel attention mechanism, referred to as Parallel Attention Super-Resolution Generative Adversarial Network (PASRGAN). The proposed algorithm adopts parallel execution of channel and spatial attention mechanisms to overcome the information dispersion issues inherent in traditional attention mechanisms. Furthermore, a feature grouping and channel shuffle strategy is introduced, which enhances feature diversity while maintaining low computational costs, thereby significantly improving the reconstruction performance of pathological images. Considering that most existing super-resolution studies on pathological images are conducted on simulated datasets that fail to fully reflect the challenges of real-world image degradation, this study constructs paired Low-Resolution (LR) and High-Resolution (HR) image datasets based on the Camelyon16 dataset to validate the effectiveness of the proposed algorithm in real-world scenarios. Experimental results demonstrate that, compared to state-of-the-art super-resolution methods (e.g., ESRGAN, CWT-Net, Histo-Diffusion, and URCDM), PASRGAN achieves superior performance with a Peak Signal-to-Noise Ratio (PSNR) of 25.33 dB, Structural Similarity Index Measure (SSIM) of 0.665 9, and Perceptual Index (PI) of 5.14. Additionally, PASRGAN achieves significantly lower parameter complexity (1.08×107) and Floating-Point Operations (FLOPs) (4.89×109) than traditional methods, confirming its computational efficiency. Ablation studies further analyze the contributions of the parallel attention mechanism, shuffle operation, and improvements in the generator and discriminator structures to verify their effectiveness.
-
Computer Engineering. 2026, 52(7): 199-209. https://doi.org/10.19678/j.issn.1000-3428.0070385
The Transformer model, owing to its excellent global information capture and powerful representation abilities, has been widely used in medical image segmentation and has achieved remarkable results. However, when serializing images, these methods divide them into fixed-sized blocks and extract only global features of a single scale, which, to some extent, fragments the semantic features of the images, ultimately leading to poor segmentation accuracy. To address this issue, this paper proposes a Multi-scale self-attention Transformer architecture (MultiFormer). First, continuous convolution and downsampling modules are used to process images. Then, the original 1×1 projection module is replaced with a multi-scale convolution projection module. Finally, deformable convolution is introduced into the feature maps generated by the self-attention module. Compared to the traditional Transformer image serialization process, this continuous convolution effectively enlarges the receptive field while generating features of the same resolution, retains the spatial correlation of 2-Dimensional (2D) images, and avoids the loss of semantic information caused by fixed-position and fixed-size image blocks. Meanwhile, the multi-scale convolution projection module captures contextual information in the images using four convolution kernels of different sizes and achieves multi-scale feature fusion through channel concatenation, reflecting attention between local regions at different scales, rather than being limited to a single scale, which enables the model to aggregate semantic information at different scales and further alleviates the problem of semantic fragmentation. Deformable convolution introduces an additional convolution layer to learn and generate an offset field, allowing the convolution kernel to flexibly adjust its shape to adapt to morphologically diverse lesions or organs in the images and enhancing the ability of the model to process complex medical images. This module is inserted into three network structures—SETR, TransUNet, and TransFuse—and experiments are conducted on the ACDC heart and ISIC 2018 skin lesion datasets. The results show that the Dice coefficient improves by 3.63, 1.06, and 2.30 percentage points and by 1.22, 2.31, and 3.01 percentage points, respectively. MultiFormer is a plug-and-play tool and can be easily integrated into various downstream medical image analysis tasks.
-
Computer Engineering. 2026, 52(7): 210-219. https://doi.org/10.19678/j.issn.1000-3428.0070747
Real-time semantic segmentation, one of the core tasks in computer vision, plays a crucial role in various applications such as autonomous driving and traffic control. Existing real-time semantic segmentation algorithms based on the encoder—decoder structure often sacrifice segmentation accuracy to achieve real-time performance. However, to ensure real-time performance, these algorithms typically have a small receptive field, resulting in poor segmentation performance for large-scale objects in road scenes. To address this issue, this paper proposes a real-time semantic segmentation algorithm tailored for road scenes based on the encoder—decoder structure. First, a Multi-scale Feature Fusion (MFF) mechanism is designed in the feature extraction stage to effectively fuse receptive field features on a larger scale, thereby enhancing the segmentation performance for large-scale objects. Second, a Polarized Self-Attention (PSA) mechanism is incorporated at the end of the encoder to enhance the local perception within the large-scale receptive field, further improving the segmentation performance for large-scale objects. Experimental results on the Cityscapes and Camvid datasets show that, when using a single NVIDIA RTX 3090 GPU, the algorithm achieves Mean Intersection over Union (MIoU) scores of 80.6% and 81.1% at frame rates of 43.5 frame/s and 91.2 frame/s, respectively. These results show that the proposed algorithm achieves higher segmentation accuracy than the comparative algorithm.
-
Computer Engineering. 2026, 52(7): 220-232. https://doi.org/10.19678/j.issn.1000-3428.0070531
Improving dense pedestrian detection performance is a key issue in the development of crowd monitoring systems for large public places. Existing systems present difficulties in detecting small targets due to crowd occlusion in dense pedestrian detection scenarios and do not adequately meet the deployment requirements for lightweight models. To address these issues, this paper proposes an improved dense pedestrian detection model, CAD-YOLO, based on YOLOv8-n. Introducing a CGDown downsampling module and an efficient context information extraction mechanism helps in effectively mitigating the problem of context feature loss, which is encountered by traditional object detectors when dealing with dense scenes. These inclusions significantly enhances the model's ability to capture dense pedestrian features and focus on small targets. A BiFPN-Adaptive structure is designed, and the neck network is reconstructed, allowing the model to more accurately extract the features of occluded pedestrians as well as small- and medium-sized pedestrians by adaptively fusing feature information at different scales. Further, the introduction of the dynamic detection head Dyhead, combined with a newly added small-target detection layer with a size of 160×160, enables the model to more precisely capture the subtle features of dense small target areas, thereby effectively alleviating the problem of missed detections in occlusion scenarios. Experimental results show that, compared to YOLOv8-n, CAD-YOLO achieves a detection accuracy improvement of 5.1 and 2.1 percentage points on the Crowd Human and WiderPerson datasets, respectively. Additionally, CAD-YOLO has a parameter count of 2.9×106 and model computation of 12.3×109; thus, it satisfies the requirements of low power consumption and high accuracy for deployment on edge devices or mobile devices.
-
Computer Engineering. 2026, 52(7): 233-242. https://doi.org/10.19678/j.issn.1000-3428.0070592
Fusion-based RGB-D Salient Object Detection (RGB-D SOD) methods employ a dual-stream structure to balance RGB and depth maps, aiming to achieve superior performance in SOD compared with using RGB maps alone. However, owing to factors such as the scene and shooting equipment, depth maps vary in quality. Low-quality depth maps cannot effectively assist RGB images in salient detection and may even have a negative impact. Existing methods rarely consider the quality differences in depth channels, leading to the unreasonable fusion of interference information from low-quality depth regions, which affects the accuracy of the results. To address these issues, this paper proposes a fine-grained depth-quality-aware RGB-D SOD method, which is divided into two stages: fine-grained depth quality awareness and bilinear routing fusion. In the first stage, the depth channel quality is perceived by comparing the similarity differences between the salient results of the RGB and D branches and Pseudo Ground Truth (PGT) in a fine-grained manner. In the second stage, a new bilinear routing fusion strategy is designed based on the quality perception results for different depth regions. The routing direction is driven by the results of depth channel quality perception, achieving more effective multimodal complementary results than existing fusion schemes. The proposed method is compared to nine State-Of-The-Art (SOTA) methods on the NJU2K, NLPR, SIP, and STERE datasets. The results show that the proposed method outperforms current leading models in all evaluation metrics. Compared with the suboptimal models, the S-measure and F-measure metrics of the proposed method are improved by an average of 0.6% and 0.7%, respectively, and the Mean Absolute Error (MAE) is reduced by an average of 0.4%. The proposed method significantly reduces the impact of low-quality, non-contributing, or negatively impacting depth regions on the results, achieving better multimodal fusion and enhancing the effectiveness of salient detection.
-
Computer Engineering. 2026, 52(7): 243-254. https://doi.org/10.19678/j.issn.1000-3428.0070434
Skin lesion images significantly affect the accuracy of segmentation algorithms because of characteristics such as blurred lesion boundaries, large variations in size, and irregular shapes. To address this issue, a skin lesion image segmentation network that combines feature learning and boundary guidance is proposed. First, a dual feature encoding module based on Convolutional Neural Networks (CNN) and a Transformer is constructed to learn the global-local features of skin lesion images. Subsequently, expert-annotated ground truth supervised learning is adopted to guide the generation of the skin lesion boundary windows and enhance the boundary features of the lesion. Finally, by introducing a boundary-guided decoding mechanism, the learned global-local features and boundary-enhanced features are fused to perform skin lesion image segmentation, optimizing the boundary clarity and overall accuracy of the segmentation results. Experimental results on the ISIC 2018 and ISIC 2016 datasets show that the proposed method achieves improvements of 2.88, 1.25, and 0.27 percentage points in Intersection over Union (IoU), specificity, and accuracy, respectively, compared with CTO on the ISIC 2018 dataset. For the ISIC 2016 dataset, the Dice coefficient and IoU are improved by 1.66 and 6.02 percentage points, respectively, achieving optimal results across all evaluation metrics. The proposed segmentation network exhibits excellent performance in segmentation tasks with complex backgrounds and blurred boundaries, accurately segmenting lesion regions and effectively handling boundary blurring issues in complex backgrounds, thereby providing a high-precision solution for skin lesion image segmentation.
-
Computer Engineering. 2026, 52(7): 255-264. https://doi.org/10.19678/j.issn.1000-3428.0070282
Existing digital signature schemes are widely used in e-commerce, financial transactions, and medical systems. However, they cannot modify the content of signed messages while keeping the signature unchanged. Moreover, repeatedly performing signature operations increases complexity and latency. To address this issue, a digital signature scheme based on the SM2 algorithm and a supporting threshold modification, called SM2-TM, is proposed. This scheme supports the collaborative modification of signed message content in a threshold manner and achieves signature immutability through parameter control and the construction of hash collisions. Furthermore, it effectively solves complex certificate storage and management as well as key escrow issues. Privacy and security analyses results show that the SM2-TM scheme possesses unforgeability, immutability, and traceability. The performance analysis results indicate that the SM2-TM scheme has certain advantages in terms of computational overhead and significant advantages in terms of the storage overhead of public keys, private keys, and signatures, with storage overheads of only 64, 32, and 64 Byte, respectively.
-
Computer Engineering. 2026, 52(7): 265-276. https://doi.org/10.19678/j.issn.1000-3428.0070708
To address the contradiction between privacy protection and recognition performance in face recognition technology, a face privacy protection and recognition method combining generative and embedding networks is proposed. First, a Convolutional Neural Network (CNN)-based generative model is employed to randomly perturb the pixel values of face images, generating distorted face images that are unrecognizable to human vision but can be identified by specific deep neural networks, thereby forming cancelable face templates. Subsequently, features are extracted using a pretrained FaceNet embedding network model for recognition. During the training process, to ensure the recognition of face templates, a residual structure is utilized in the generative network model to effectively extract key features from the original images, enhance the expressiveness of the images, and reduce information loss to a certain extent. To increase the differences between the original and generated images and improve the diversity of the generated images, generative hybrid loss and diversity loss functions are introduced. To enhance the recognition accuracy, an improved triplet loss function is employed to optimize the model. The research results demonstrate that this method not only improves the privacy security of face templates but also enhances the differences among generated images using the diversity loss function, thereby improving model robustness. Experimental results on the Aberdeen, GT, and LFW datasets indicate that the improved triplet loss function achieves more representative feature representations in the cosine-embedding space, with recognition accuracy rates of 99.87%, 99.29%, and 98.59%, respectively.
-
Computer Engineering. 2026, 52(7): 277-286. https://doi.org/10.19678/j.issn.1000-3428.0070542
The development of quantum computing threatens the security of existing public-key cryptosystems. To prevent the ″store-now-decrypt-later″ attack, existing public-key cryptosystems must urgently migrate to post-quantum cryptosystems. The Aigis-sig digital signature scheme and Aigis-enc key encapsulation scheme, constructed based on ideal lattices, are the first-prize winners in China's cryptographic algorithm design competition and possess advantages in terms of resisting quantum attacks. The codes of the two systems are integrated to enhance resource utilization and enable the efficient application of Aigis-sig and Aigis-enc with limited hardware resources. Two sets of butterfly operators are designed on the hardware modules, significantly improving the computational efficiency of the fast Number Theory Transformation (NTT) through pipelined operations. Building on this, a hardware—software co-design implementation method for the Aigis-sig and Aigis-enc schemes is proposed. Experimental tests demonstrate that this approach achieves notable performance improvements compared to a pure software implementation. Specifically, the ROM space usage is reduced by 65%, the average execution times for digital signing and verification are shortened by 29% and 11%, respectively, and the average execution times for key encapsulation and decapsulation are shortened by 13% and 21%, respectively.
-
Computer Engineering. 2026, 52(7): 287-301. https://doi.org/10.19678/j.issn.1000-3428.0070755
Deep neural network language models are vulnerable to adversarial example attacks during application. To address this issue, adversarial examples are typically generated by adding minor perturbations to the original samples to mislead the model into making incorrect decisions. By studying generation methods for adversarial examples, the robustness of the models can be evaluated. Most existing Chinese adversarial example generation methods focus on improving the attack success rate of adversarial examples while overlooking quality indicators such as their stealthiness. Therefore, focusing on Chinese text adversarial example generation techniques and incorporating the unique characteristics of Chinese in terms of glyph structure and semantic features, a multi-level adversarial example generation method called CMSPSO is proposed that combines Chinese character glyphs and semantic information. The CMSPSO utilizes the Particle Swarm Optimization (PSO) algorithm to search for appropriate replacement combinations in a predesigned replacement knowledge base to generate adversarial examples. CMSPSO-M combines the multilingual visually similar character features of Chinese characters, calculates the similarity of visually similar characters through a trained Siamese Neural Network (SNN), and constructs a high-quality visual replacement character knowledge base to generate character-level adversarial examples. CMSPSO-S constructs a semantic replacement word knowledge base based on HowNet and WordNet to generate word-level adversarial examples and evaluates them using attack effectiveness and attack cost metrics. The experimental results demonstrate that CMSPSO exhibits significant attack effectiveness across multiple models and datasets, particularly on the XNLI dataset, where CMSPSO-M achieves an attack success rate of 84.22% against the RoBERTa model. Additionally, CMSPSO shows distinct advantages in terms of attack cost metrics, outperforming the baseline methods in terms of overall performance.
-
Computer Engineering. 2026, 52(7): 302-312. https://doi.org/10.19678/j.issn.1000-3428.0070375
As Internet platforms progress rapidly, recommendation systems are challenged by degraded accuracy due to shilling attacks, while still providing personalized services. Existing shilling attack detection algorithms mostly focus on a single or a few evaluation metrics from the perspective of user rating differences and seldom consider the preference correlation of the items selected by users. Consequently, user behavior characteristics are modeled insufficiently, resulting in high misdetection rates or limited applicability to different attack patterns. To address this issue, this paper proposes a Multi-View Feature Fusion Shilling Attack Detection (MVFF-SAD) algorithm. Based on the latent features and distribution characteristics of user ratings obtained through a Variational AutoEncoder (VAE), the algorithm learns the spatiotemporal distribution characteristics of user profiles from the perspectives of short- and long-term dependencies of ratings and probability density distributions. By combining the historical preference correlation of users, the algorithm uses a neural network model for multi-view feature fusion to form a comprehensive user profile representation with enhanced detection capability, thereby improving shilling attack detection accuracy. Experimental results show that the proposed algorithm significantly improves the detection accuracy for fake users, achieving an accuracy of over 95% in most cases. The proposed algorithm also demonstrates good detection performance under different filling rates and attack scales, and exhibits strong robustness.
-
Computer Engineering. 2026, 52(7): 313-330. https://doi.org/10.19678/j.issn.1000-3428.0070476
Deep Neural Networks (DNNs) are vulnerable to adversarial examples and adding even a small perturbation to a clean image can cause a misclassification. Decision-based attacks are a class of black-box attacks that rely only on a target model to predict hard-label outputs. They consider the target model as a black box, and the attack queries the results of the target model without requiring access to its internal structure or parameter information. This poses a significant threat in real-world applications. Current decision-based attacks typically utilize gradient estimation to launch attacks near the decision boundary of the target model; however, they require high query costs and generate poor-quality adversarial examples with more severe distortion. The results of this study show that low-frequency information in the frequency space of an image can effectively capture important features of an image. A decision-based attack in a low-frequency space not only helps reduce the number of queries but also generates high-quality adversarial samples. To this end, this study proposes a black-box attack method based on the geometric properties of circles, called CBA. This method utilizes a Discrete Cosine Transform (DCT) to obtain adversarial examples in the frequency space using the geometric properties of circles near the decision boundary in a continuous iteration. Finally, the Inverse Discrete Cosine Transform (IDCT) changes them back into input space. This avoids gradient estimation and significantly reduces the number of queries while guaranteeing the success rate of the attack. Experimental results on the ImageNet dataset show that the attack success rate of CBA for generating adversarial examples is higher than that of the latest black-box attack methods, which utilize the geometric nature of the decision boundaries for query volumes of 500, 1 000, 2 000, respectively. Additionally, CBA has a higher attack success rate under different constraints for the same query volume. These results show that CBA reduces the number of queries required to generate adversarial examples and generates adversarial examples with less distortion and better image quality. Additionally, the effectiveness of CBA is tested using a real-world model.
-
Computer Engineering. 2026, 52(7): 331-344. https://doi.org/10.19678/j.issn.1000-3428.0070469
Scene elements are fundamental for understanding urban geographic information, and their accurate extraction is essential for smart city development and geographic information systems. To address the complexity of street-view images, the limitations of existing deep learning models in interpreting complex scenes, and challenges in associating visual data with context, a method based on large multimodal models for extracting typical scene elements from street-view images is proposed. First, the approach extends the Large Language and Vision Assistant (LLaVA) by integrating a multilayer perceptron and a high-resolution visual encoder to create GeoLLaVA. Second, a Street View Visual-Instruction-Following Dataset is constructed for scene element extraction tasks to provide multidimensional instructions. The model is fine-tuned using visual instructions to enhance contextual understanding. Low-Rank Adaptation (LoRA) is used to optimize the computational efficiency. Finally, GeoLLaVA generates multidimensional scene descriptions from street-view images and extracts key element keywords for effective scene-element extraction. In comparative experiments using semantic segmentation, object detection, and other multimodal models, GeoLLaVA demonstrates significant advantages, achieving F1 scores of 0.938, 0.842, and 0.829 for the extraction of traffic signals, intersections, and parking lots, respectively. A comparison between the model before and after fine-tuning clearly demonstrates the effectiveness of the fine-tuning process. Ablation studies further validate the performance improvements achieved by the modified GeoLLaVA architecture, and LoRA effectively reduces computational resource consumption. Regional application experiments using batch inference on street view images with geographic coordinates and a comparison with OpenStreetMap (OSM) data not only confirm the accuracy of the model but also highlight the limitations of OSM data in providing comprehensive element information.
-
Computer Engineering. 2026, 52(7): 345-353. https://doi.org/10.19678/j.issn.1000-3428.0069927
To overcome the limitations of traditional single-modal Named Entity Recognition (NER) methods in processing multimodal data, this study proposes a novel cross-modal NER method based on semantic fusion. Unlike conventional approaches that use pretrained language models combined with Conditional Random Fields (CRFs), the proposed method introduces image information into the NER task and integrates the semantic content conveyed by images. First, the research status of NER tasks is reviewed, followed by a detailed explanation of the two core modules of the proposed cross-modal NER method: the image—text semantic fusion module and the bidirectional cross-modal matching module. Specifically, the image—text semantic fusion module generates textual representations based on the semantic content reflected in the images corresponding to the text, which are then combined with the original text information to form new text sequences. The bidirectional cross-modal matching module utilizes the newly generated text sequences to perform cross-modal matching with images. This approach of jointly considering image semantics and text before conducting cross-modal matching not only improves the accuracy of NER but also enriches the semantic representations of named entities. Finally, the effectiveness and superiority of the proposed method are demonstrated through comprehensive comparative and ablation experiments on standard benchmarks (Twitter-2015 and Twitter-2017). This study provides new insights and presents an effective methodological advancement for the NER of multimodal data.
-
Computer Engineering. 2026, 52(7): 354-367. https://doi.org/10.19678/j.issn.1000-3428.0252238
In response to the challenges in existing Weakly-supervised Temporal Action Localization (WTAL) research, such as the underutilization of action temporal characteristics, global properties, and action semantic consistency, a method called Weakly-Supervised Action Localization via Video—Text Semantic alignment and Full Video Dependency (FVD-ALM) is proposed. First, dilated convolution networks are used to expand the receptive field of the model, and attention mechanisms are utilized to precisely enhance the temporal features of action instances, ensuring accurate temporal feature extraction. Subsequently, an Expectation-Maximization (EM) algorithm based on a Gaussian Mixture Model (GMM) is applied to extract and enhance global information from the video, generating accurate temporal class activation maps to aid in the action localization process. Finally, a video—text semantic alignment module is designed to comprehensively understand actions by combining textual information with action labels. The model is trained to complete textual descriptions of actions, thereby enhancing its cognitive ability for action-category consistency and effectively distinguishing different action categories. Experimental results on the THUMOS14 and ActivityNet1.3 datasets confirm the effectiveness of this method, achieving an average mean Average Precision (mAP) of 39.1% on THUMOS14, which is a 2.0 percentage points improvement over the DTRP-Loc method. This demonstrates that the method of integrating multisource information significantly improves the accuracy of action localization and provides an effective solution for weakly-supervised action localization tasks.
-
Computer Engineering. 2026, 52(7): 368-376. https://doi.org/10.19678/j.issn.1000-3428.0070493
The Multimodal Relation Extraction (MRE) method assists in relation extraction tasks by using multimodal information. To achieve a good relation extraction performance, existing MRE models usually need to be effectively trained on a large amount of labeled data. However, they perform poorly in the case of few-shot samples. To address this issue, this study utilizes the rich semantic and prior knowledge of relation labels and proposes a new prompt-tuning hierarchical network. First, a text-prompt module based on knowledge injection is designed and implemented. Specifically, by utilizing the entity information hidden in relation labels and introducing virtual entity type words to construct prompt templates, the model can perceive the potential range of entity types in the sample and continuously optimize the introduced virtual relation answer words using context to express the best semantic information, thereby improving the performance of the model in the case of few-shot samples. Second, by utilizing the mutual constraint relationship between entity pairs and relations, an entity—relation collaborative optimization module is designed and implemented to further improve the effect of relation extraction. Finally, in each self-attention layer of the text encoder, a visual prefix-based attention mechanism is introduced to deeply integrate the layered multiscale visual features with the text information, thereby generating more effective and robust text representations and significantly reducing the sensitivity of the model to errors. The experimental results on the multimodal neural relation extraction dataset (MNRE) show that the precision, recall, and F1 score of the model reach 84.97%, 83.91%, and 84.43%, respectively, which are all improved compared with the benchmark model. Especially in the case of few-shot samples, the model proposed in this study is significantly better than the benchmark model, demonstrating good relation extraction performance.
-
Computer Engineering. 2026, 52(7): 377-389. https://doi.org/10.19678/j.issn.1000-3428.0070654
With the increasing energy consumption of urban rail transit systems, enhancing the utilization of regenerative braking energy to reduce the energy consumption of train operations has become a critical issue. This study focuses on the optimization problem of tracking train operation control strategies in the multi-train cooperative operation process. First, building upon the traditional transition strategy based on operation mode, the study proposes the ″Traction—Coasting—Traction—Cruising—Coasting—Braking (TCTCCB)″ strategy for the tracking operation scenario. Second, the study constructs a train dynamics model in the spatial domain, a state transition equation, and an energy consumption model. By employing the interpolation method, the cooperative operation problem in the time domain is transformed into a problem of solving optimal switch points in the spatial domain. Subsequently, an optimization decision-making model with the goals of energy consumption and punctuality is constructed, which is then efficiently solved using the Dung-Beetle Optimizer (DBO). Finally, considering the Yizhuang Line of the Beijing Subway as a simulation line, comparative analyses are conducted to evaluate the influence of Communication-Based Train Control (CBTC) and Train Autonomous Control System (TACS) architectures, as well as different transition strategies, on optimization performance. The results demonstrate that TACS significantly enhances the optimization performance of cooperative operations compared with CBTC. The proposed strategy not only meets the punctuality requirement but also outperforms the traditional strategy in terms of energy consumption at various departure intervals. The net absorbed energy consumption increases by a maximum of 14.651 kWh, and the actual operational energy consumption decreases by a maximum of 11.284 kWh. Therefore, the proposed operational mode transition strategy and optimization method effectively improve the energy consumption of train operations and have reference significance for the development of urban rail train operation control technologies.
-
Computer Engineering. 2026, 52(7): 390-402. https://doi.org/10.19678/j.issn.1000-3428.0070583
With the rapid development of autonomous driving and smart cities, accurate road point cloud segmentation and reconstruction have become key technological requirements. Traditional segmentation methods face difficulties when dealing with complex road geometries. To address this issue and maximize the advantages of oblique photography technology to quickly generate a real-field scenic spot cloud model, this study proposes a new segmentation optimization method. This method introduces an automatic optimization strategy guided by the connectivity evaluation index and combines a local coding module based on the normal vector, curvature, and other features. Consequently, the ability of the semantic segmentation network to capture complex road geometric features is enhanced, thereby improving the segmentation accuracy and enhancing the overall continuity of road structures. In addition, the study proposes an iterative segmentation optimization strategy based on road median line extraction and repair. On the one hand, the problem of compensating unavoidable hollow point clouds in oblique photographic point clouds is transformed into easy problems such as road median line identification and repair and road width estimation, and on this basis, the existing road median line extraction algorithm is optimized. The problems of road deformation and incompleteness are solved using traditional algorithms. On the other hand, the road median and road width are innovatively taken as prior knowledge to assist the point cloud semantic segmentation model in extracting roads with higher accuracy. Finally, a complete set of procedurally generated road-flow schemes is implemented, which can be used to automate and stably restore complex road models with different types of intersections. Experimental results verify that the proposed method significantly improves the reconstruction effect, providing a new concept and application value for the progress of road point cloud processing technology.
-
Computer Engineering. 2026, 52(7): 403-410. https://doi.org/10.19678/j.issn.1000-3428.0070484
Airspace traffic complexity is an important factor affecting the efficiency and safety of civil aviation operations. To further improve the accuracy of airspace traffic complexity assessments, this study proposes a complexity evaluation method based on multi-scale airspace traffic spatio-temporal images and deep metric learning. Specifically, traffic flow data are mapped to grid-based target airspace images in the form of pixels. Spatio-temporal interpolation is performed to capture the dynamic changes in traffic flow over time and space, resulting in the generation of 20 sets of spatio-temporal airspace traffic images at different scales. Then, an airspace traffic complexity assessment model based on deep metric learning is proposed, which takes multi-scaled airspace traffic image sets as the input. The model uses a ranking proxy anchor loss function to optimize the distribution of the sample distances in the high-dimensional embedding space, such that the same-class samples are closer and the different-class samples are more dispersed. Experiments are conducted using real traffic data from the south-central airspace to generate a multi-scale spatio-temporal image set, followed by a series of comparative experiments. The results show that the spatio-temporal scale of the airspace traffic image sets has considerable impact on the assessment results. Compared with existing assessment methods, the proposed method can significantly improve the assessment performance of airspace traffic complexity.
-
Computer Engineering. 2026, 52(7): 411-420. https://doi.org/10.19678/j.issn.1000-3428.0070351
During pneumoconiosis staging, the ambiguity in staging criteria leads to varying diagnostic results, causing instance label noise where label accuracy is affected by observer variability. Training deep learning models on such noisy labeled data often causes performance degradation. To address the challenges posed by noisy labels in pneumoconiosis staging, this study proposes a dual-branch dynamic sample selection strategy. In the key layer of the model, the method introduces Monte Carlo Dropout (MC-Dropout) and performs multiple forward propagation for uncertainty estimation of the prediction results, while combining the prediction probability of the model with the Gaussian Mixture Model (GMM) to gain insights into the feature spatial distribution of the samples. It ultimately realizes the branches of selective uncertainty estimation and Predicting the Feature Distributions of the GMM (FDP-GMM). Subsequently, a threshold value is set for each sample data in each branch, which is momentum updated with the model training. Finally, the dynamic sample selection module is used to divide the samples and differentiate the constraints to reduce the negative impact of noisy samples on the model. The method is evaluated on a clinically collected pneumoconiosis X-ray dataset, and the results show that it significantly outperforms similar methods at noise rates of 5%, 10%, 20%, and 40%, demonstrating its effectiveness and robustness in learning from noisy labels.
-
Computer Engineering. 2026, 52(7): 421-433. https://doi.org/10.19678/j.issn.1000-3428.0070379
As export container throughput continues to increase in container terminal yards, ensuring the efficient management and scheduling of terminal resources has become crucial for enhancing terminal competitiveness. Considering an automated container terminal yard as the research object, this study proposes an optimization method based on Deep Reinforcement Learning (DRL) to address the issues of storage slot allocation and yard crane scheduling for export containers. A comprehensive analysis of the container terminal yard operation system is performed, and a multi-objective optimization model is constructed. This model primarily aims to reduce the yard crane operation time, internal truck waiting time, and number of container relocations within the yard blocks, while also considering constraints such as the safe distance between yard cranes and balanced workload distribution. To reduce the complexity of solving the model, an Actor—Critic algorithm based on DRL is proposed. Numerical examples of different scales are designed. Through a comparative analysis with the results of the Genetic Algorithm (GA) and exact solutions from CPLEX, the advantages of the Actor—Critic algorithm in terms of solution speed and quality are demonstrated. Experiments reveal that the proposed algorithm can quickly and accurately solve small-scale problems and obtain near-optimal solutions for large-scale problems, exhibiting significant superiority over the GA in solving large-scale problems. Comparative experiments are conducted to further investigate the impact of the number of yard cranes and zoning-balanced mixed stacking strategy on the optimization results. The analysis results indicate that the zoning-balanced mixed stacking strategy outperforms traditional stacking strategies in terms of balancing yard crane workloads and reducing idle time.
-
Computer Engineering. 2026, 52(7): 434-444. https://doi.org/10.19678/j.issn.1000-3428.0070391
Structured medical data are widely used in real-world healthcare systems. Owing to their unique characteristics, they typically exhibit properties such as being multi-source and heterogeneous with many potential candidate features, resulting in irregular data patterns that pose challenges to the feature-processing procedures of neural networks. Therefore, a recurrent directional feature processing network with a feature-embedding level similar to that of the Transformer architecture, named RDFP-NET, is proposed. The network possesses a favorable feature organization pattern and performs end-to-end processing directly on the collected raw medical data, enabling it to fully leverage effective information while eliminating the negative impact of unfavorable features. In addition, it employs a recurrent structure to correlate contextual feature information, further enhancing network task performance. RDFP-NET mainly comprises: a nonlinear feature-embedding method; a ranked feature interaction module that incorporates inductive bias into the general attention mechanism by calculating the information value of each feature and applying specific masking operations, achieving directional information flow and feature interaction among features with varying contribution values; a filtered feature selection module that discards redundant features in a filter-like manner by adjusting the Exponential Linear Unit (ELU) activation function; and a recurrent structure comprising two modules that facilitates good data fitting and contextual feature information flow while gradually addressing irregular feature patterns. Experiments conducted on five datasets encompassing two different target tasks demonstrate that RDFP-NET achieves an optimal average rank of 2.2 and a rank standard deviation of 0.75, showing excellent task performance and stability. Thus, RDFP-NET can provide medical professionals with more effective auxiliary information for clinical decision-making.
Volume 52 No 7
Superintended by:
China Electronics Technology Group Corporation
Sponsored by：
East-China Institute of Computer Technology
Shanghai Computer Society
Edited & Published by：
Editorial Office of Computer Engineering
Chief Editor：Wu Jiangxing
Executive Chief Editor:
Hu Bing
Tel：021-67092514
E-mail: ecice06@ecict.com.cn
CN: 31-1289/TP
ISSN: 1000-3428