RGN-UOD: An information redistribution-guided network for underwater object detection
Abstract
Underwater object detection (UOD) plays a vital role in marine resource exploration, underwater ecological monitoring, military security, and related fields. However, underwater image degradation, such as color casts, haze, and low contrast, can easily lead to target feature drift. Moreover, most traditional image enhancement methods involve irreversible operations, which lead to the loss of original information and severely limit detection accuracy. Existing underwater detection methods either fail to specifically address the interference of image degradation or suffer from limited performance gains due to the lack of tight coupling between enhancement and detection, making them difficult to adapt to complex underwater scenarios. To this end, this paper proposes an end-to-end underwater object detection network guided by information redistribution, named RGN-UOD. The network introduces a lightweight information redistribution module at the front end of the detector. Inspired by invertible transformations, the module maps degraded inputs to a representation space that is more favorable for detection through image-adaptive channel redistribution, thereby alleviating feature drift in underwater scenes while preserving the recoverability of the input as much as possible. Then, a DW-FEM module is designed to enhance local target features from multiple aspects through multi-branch asymmetric convolutions, enriching texture and semantic representations. Finally, an improved hybrid encoder is constructed, which integrates AIFI global attention and SS2D lightweight global modeling to enable sufficient intra-scale interaction and deep cross-scale fusion, improving target identification and localization in complex backgrounds. Experimental results on three mainstream underwater datasets, URPC2018, UTDAC2020, and DUO, show that RGN-UOD achieves mAP50 of 81.2%, 86.4%, and 86.6%, and mAP50:95 of 48.7%, 51.5%, and 67.5%, respectively, outperforming other mainstream competing methods. Meanwhile, the lightweight version has only 9.18M parameters and an inference speed of 100.8 FPS, indicating that the proposed method achieves high detection accuracy with low model complexity and fast inference.