Abstract: Globally escalating marine ecological pollution necessitates efficient and precise automated intelligent monitoring and cleaning technologies for underwater litter leveraging autonomous underwater vehicles (AUVs) or underwater robots, which has become an urgent academic consensus in the field of marine environmental protection. To address the inherent physical degradation of underwater visible light imaging—such as color casts, blurring, and low contrast—as well as the bottlenecks of cross-domain heterogeneity caused by varying optical properties of water bodies across different sea areas and the severe scarcity of high-quality fine-grained annotated samples, this study aims to overcome the feature shift and negative transfer phenomena induced by deploying models in unknown water environments. To this end, it proposes an underwater litter object detection framework that merges pixel-level image enhancement with a two-stage robust cross-domain adaptive learning scheme. This framework adopts a design strategy of "step-by-step construction and dynamic alignment," decoupling the pipeline construction into two mutually collaborative training stages. The first stage aims to construct an interference-resistant baseline perception architecture, establishing a solid feature foundation of deep representations for the entire network. To maintain high computational efficiency while circumventing the limitations of complex deep generative networks—which easily introduce spurious edges and grid-like artifacts under small-sample conditions—a low-cost pixel-level physical alignment pipeline is implemented at the data loading layer. Specifically, the input blue-green (BGR) raw images are first transformed into the LAB color space, where the L channel is extracted and subjected to adaptive pixel-level linear stretching to maximize the restoration of edge contrast and fine details. Subsequently, a Channel Stabilization Module compensates for color casts in the A and B channels to robustly suppress underwater-specific cyan-blue artifacts, before inversely mapping the calibrated feature maps back to the standard red-green-blue (RGB) space. On this basis, 30,000 iterations of supervised perceptual training are executed on the source dataset (TrashCan Instance), enabling the RetinaNet detector to acquire high-quality low-level general geometric textures and high-level robust semantic features, thereby establishing a solid perception baseline for multi-source scenarios. The workflow then naturally proceeds to the second stage, which focuses on domain alignment and precision enhancement for cross-domain deployment. To eliminate deep feature shift and response distortion triggered by unknown new water domains, a lightweight Feature Calibration Module (FCM) is symmetrically injected after the multi-scale output layers of the pre-trained backbone network. A differential initial assignment based on the Kaiming normal distribution is directly applied to the internal parameters of the FCM at the very beginning of the fine-tuning process to explicitly induce adaptive dynamic calibration. During the subsequent 10,000 iterations of cross-domain adaptive fine-tuning, to avoid catastrophic forgetting under low-data regimes, this stage completely freezes the feature extractor parameters of the backbone network, serving as a locked low-level feature stabilizer. Concurrently, a differential asymmetrical dual-learning-rate mechanism is deployed, assigning high gradient sensitivity (initial learning rate 0.01) to the newly injected FCM while allocating a smaller step size (initial learning rate 0.002) to the top-level Feature Pyramid Network (FPN) and detection head. Furthermore, an Exponential Moving Average (EMA) technique with a fixed decay rate (λ = 0.9999) is introduced throughout the entire fine-tuning process. It automatically switches to the smoothed EMA weights during the single complete evaluation performed every 1,000 iterations, leveraging a smooth weight loss plane to shield the model from training noise brought by extremely small samples and ensuring a lossless migration of feature space alignment. To guarantee the scientific rigor of the experimental design and its practical deployment reference value, this study introduces completely independent third-party cross-domain underwater datasets for blind generalization validation. Quantitative results demonstrate that while maintaining a lightweight and compact architecture with only 55.86 million parameters, the proposed method achieves a real-time inference speed of up to 44.61 FPS on a single NVIDIA RTX 4080 SUPER GPU, showcasing immense potential for edge-side practical deployment. In in-domain closed-loop testing, the framework significantly improves the bounding box Average Precision (AP) by 5.33 percentage points compared to the traditional RetinaNet baseline model, with a notable jump of 8.97 percentage points in the highly challenging task of detecting small targets characterized by blurred edges and low visibility underwater. In the most rigorous cross-water and cross-dataset generalization blind tests, the proposed method demonstrates remarkable performance gains: on the completely unseen independent J_Litter dataset, the AP_50 of the baseline model is as low as 0.31% due to multi-source interference, whereas the proposed method catapults the AP_50 to 28.66% through dynamic feature calibration; on the DeepPlastic deep underwater dataset, the method successfully elevates the baseline AP_50 from 3.30% to a stable 12.81%. These outcomes robustly validate the scientific validity and synergetic efficiency of the "physical foundation construction" and "dynamic feature calibration" within the two-stage architecture, proving its capability to effectively overcome negative transfer in cross-water detection and highlighting its strong practical generalization potential for real-world engineering deployment in automated marine litter removal.

摘要： 当今全球海洋生态环境污染日益加剧，塑料、金属等固体废弃物在水下各深度海域激增，对海洋生物链与生态系统安全构成了前所未有的威胁。依托自主无人水下航行器或水下机器人实现高效、精准的自动化智能垃圾监测与清洁技术，已成为海洋环境保护领域的迫切学术共识。针对水下可见光成像普遍存在的色偏、模糊、低对比度等物理退化问题，以及不同海域间水体光学特性引发的跨域异质性与高质量细粒度标注样本稀缺的瓶颈，为了克服未知部署水域导致的特征偏移与负迁移现象，提出一种融合物理级图像增强与两阶段鲁棒跨域自适应学习的水下垃圾目标检测框架。该框架采取“分步构建、动态对齐”的设计策略，将流水线解耦为两个协同阶段。第一阶段旨在构建抗干扰的基础检测架构以夯实特征地基，为保持高计算效率并规避生成网络在少样本下引入伪影的缺陷，在数据载入层建立低开销的像素级物理对齐流水线：先将输入图像转换至LAB色彩空间，提取L通道执行自适应像素级线性拉伸以恢复边缘对比度与细节，再通过通道稳定化模块对A、B通道进行色偏补偿以抑制青蓝色伪影；在此基础上，基于源域数据集TrashCan Instance实施30k次迭代的监督感知训练，使RetinaNet获取高质量的底层几何纹理与高层鲁棒语义特征，确立感知基线。随后流程进入第二阶段，聚焦于域对齐与跨域部署精度提升。为消除未知新水域引发的深层特征流溢与响应偏移，研究在预训练主干网络多尺度输出层级联注入轻量级特征校准模块（FCM），并在微调初始阶段对FCM内部参数直接实施差异化初值赋能以显式诱导自适应动态校准。在随后的10k次跨域自适应迭代微调中，为杜绝少样本部署时的灾难性遗忘，本阶段完全冻结主干网特征提取器参数作为底层特征稳定器；同时采用差异化非对称双学习率机制，为新注入的FCM赋予高梯度敏感度，而为顶层FPN与检测头赋予较小步长。此外，微调全程引入固定衰减率（λ= 0.9999）的指数移动平均（EMA）技术，在每1000次迭代评估时自动切换至EMA平滑权重，利用平滑的损失平面抵御训练噪声，确保特征空间对齐的无损迁移。为确保实验设计的科学严谨性与实战部署的参考价值，研究引入完全独立的第三方跨域水下数据集进行盲测泛化性验证。定量结果表明，该方法在保持整体框架仅55.86 M参数量的轻量化、紧凑型设计下，在单张 NVIDIA RTX 4080 SUPER 显卡上实现了高达44.61 FPS（每秒传输帧数）的实时推理速度，展现了较高的边缘端实战部署潜力。在同域闭环测试中，该框架相比传统RetinaNet基线模型的平均精度显著提升了5.33个百分点，尤其在水下边缘模糊、低可视度的极其困难的小目标检测任务上，AP 跃升了8.97个百分点。而在最为严苛的跨水域、跨数据集泛化性盲测验证中，该方法展现出了显著的性能提升效果：在未参与训练的全新J_Litter独立数据集上，基线模型的AP_50在多源干扰下仅为0.31%，而该方法通过动态特征校准使AP_50升至28.66%；在深度水下场景的DeepPlastic数据集上，该方法成功将基线模型的AP_50从3.30% 稳步提升至12.81%。这充分验证了所提框架中“物理地基构建”与“动态特征校准”在两阶段架构下的科学性与协同效能，能有效克服跨水域检测中的负迁移现象，在海洋垃圾自动化智能清理的实际工程落地中具备较强的实战泛化潜力。