Towards lightweight coastal litter monitoring: A synergistic deep learning framework integrating efficient architectures with model compression and knowledge distillation.

The pervasive accumulation of anthropogenic litter along global coastlines poses severe threat to marine ecosystems and socio-economic activities. Traditional manual monitoring is labor-intensive and time-consuming. Deep learning based monitoring has emerged as a promising solution, yet was characterized by high computational complexity and large parameter counts, limiting practical application. We propose OceanTrashNet, an efficient framework reduces computational cost and parameter count while preserving high accuracy, which applies three components targeting the scale variation, elongated structures, and small-object nature of marine pollution: a backbone that utilizes reparameterized large-kernel convolutions for global context modeling; a neck module that employs efficient group shuffle convolutions for computationally frugal multi-scale feature fusion; and a reparameterized shared convolution detection head that combines cross-scale parameter sharing with training-time pathways for powerful feature learning. Furthermore, we apply a two-stage compression pipeline for accuracy-preserving lightweighting. The first stage, task-driven channel pruning enables interpretable removal of redundant network modules, and the second stage, adaptive learning knowledge distillation applies adaptive curriculum-based learning schedule to recover the accuracy lost. The final OceanTrashNet-Lite achieves 79.6% mAP50 with only 3.3 GFLOPs, a 46.9% computation reduction and 2.9% mAP improvement over baseline. Quantitative comparisons, statistical significance tests, and qualitative analysis via heatmaps and case studies confirm the effectiveness of the proposed design. This work provides an efficient, scalable solution for ocean trash monitoring, enabling timely clean-up operations and evidence-based environmental policy.