AFA-Underwater: Dual-Path Adaptive Feature Augmentation for Few-Shot Deep-Sea Trash Detection

Abstract

As marine pollution intensifies, developing efficient deep-sea garbage detection technologies has become increasingly urgent. However, deep-sea environments present two major challenges: (i) the scarcity of target samples (few-shot problem), which makes it difficult to train deep learning models effectively; and (ii) harsh underwater imaging conditions—such as poor illumination and scattering—that lead to object blurring and background confusion. To address these challenges, this paper proposes AFA-Underwater, a framework specifically designed for few-shot deep-sea garbage detection. The core of this framework is a dual-path Adaptive Feature Augmentation (AFA) module, which generates more robust class prototypes to effectively mitigate the “prototype shift” problem caused by underwater noise. In addition, we introduce a novel Tri-Path Aggregation (TPA) mechanism that jointly models the raw, common, and differential representations between support and query features, thereby alleviating base-class bias. Extensive experiments on the re-partitioned TrashCAN 1.0 dataset demonstrate that AFA-Underwater significantly outperforms existing FSOD methods under multiple few-shot settings, achieving a remarkable 14.4% improvement in novel-class mAP under the 10-shot scenario. The algorithm also shows strong generalization capability on the PASCAL VOC dataset.