Local and Global Feature Refinement for Underwater Object Detection

Recently, underwater object detection has attracted a lot of interest. However, due to the blur and low contrast in the underwater scene, if deployed directly, generic detection methods show poor performance. To address this problem, we propose a local and global fusion underwater object detection algorithm with a receptive field refinement module (RFRM). Specifically, LGFB combines inductive biases and long-range dependencies to obtain high-quality visual representations. This module implements the interaction between the convolution block and the transformer block. RFRM takes the different receptive fields into consideration and suppresses the redundant information by exploring the relationship between adjacent layers. Extensive experiments demonstrate that the proposed method is superior to generic object detectors, achieving 80.9 mAP
50
and 71.3 mAP
50
on URPC2018 and UDD datasets. Code is publicly available at https://github.com/caiduoduo12138/underwater.