Abstract

Plastic pollution in the ocean is a widespread issue in the marine biosphere that requires large-scale monitoring systems. However, extending the use of deep-learning-based detection methodologies to real-world marine settings is difficult. Reflections from surfaces, small target objects, and partially blocked debris often hinder the effective functioning of such systems. To mitigate these challenges, this paper presents a detection system built on the You Only Look Once architecture that explicitly considers the aforementioned constraints. The architecture combines two mutually supportive modules: a directional coordinate attention module, which decodes spatial dependencies along horizontal and vertical axes, and a Sinkhorn distance-based regularization term, which stabilizes feature focus across spatial dimensions. Experimental testing of image collections by aerial and underwater cameras shows significant performance improvements compared to the latest state-of-the-art assessments. The proposed system achieves a precision of 92% and a recall of 94% in aerial scenes and a precision of 90% and a recall of 92% in sub-aqueous scenes. An ablation study validates the hypothesis that the two modules work together to improve performance. Furthermore, visual inspection yields more reliable results for detecting typical marine debris, including reflective artifacts, small objects, and visually contaminated scenes.