SRP-UOD: Multi-Branch Hybrid Network Framework Based on Structural Re-Parameterization for Underwater Small Object Detection

Abstract

This study presents a novel multi-branch hybrid network framework based on structural re-parameterization for underwater small object detection, referred to as SRP-UOD. The developed framework employs multi-branch convolutions with ReBottleneck-ResNets, which are re-parameterized based on structural parameters, to facilitate more efficient feature extraction and fusion. Furthermore, the addition of SPD-Block enables the extraction of fine-grained information from images, thereby reducing the learning efficiency of low-quality feature representations. Specifically, we introduce a Distance-IoU Loss for the underwater target detection scenario, aiming to address the performance degradation caused by target occlusions. The resulting SRP-UOD model shows significant effectiveness in underwater target detection, achieving an average accuracy rate (mAP) of 83. 6% in the URPC2020 (Dalian) competition dataset.