Shape-Adaptive Multi-Scale Feature Enhancement Network for Object Detection in Underwater Vehicle

Abstract

Marine organisms exhibit substantial variation in appearance and shape across diverse underwater environments and are often characterized by considerably smaller object scales than their terrestrial counterparts. These properties pose significant challenges to generic object detectors, particularly in capturing shape-varying targets and preserving discriminative information for small objects. To address these underwater-specific challenges and enhance visual perception for underwater vehicles, we propose a Shape-Adaptive Multi-Scale Enhancement Network for underwater object detection, built upon YOLOv8. Our method introduces two key components: a Shape-Adaptive Downsampling Attention (SADA) module and a High-Resolution Feature Compensation (HRFC) module. The SADA module replaces the conventional downsampling operation in the neck and leverages high-level semantic features to adaptively generate attention weights for corresponding low-level features, enabling selective aggregation of shape-sensitive representations. Meanwhile, the HRFC module compensates for the loss of spatial information in low-resolution high-level features. Specifically, a scale mapper projects low-resolution features into a high-resolution feature space, where cross-scale feature fusion is subsequently performed to enhance fine-grained representations. Extensive experiments on the URPC2020 dataset demonstrate that the proposed method consistently outperforms the YOLOv8 baseline, achieving a 3.6 mAP improvement in detection performance and establishing state-of-the-art results.