Underwater Target Detection by Residual Spatial Cooperative Attention Module–Based Self-Supervised Learning

The performance of underwater target detection techniques is limited by various factors. First, in underwater image data sets, there exists degradation such as color bias, low contrast, and blurring, which affect the accuracy of the detection algorithms. Second, the underwater image data set is difficult to obtain and the cost of making the labeled data sets is high, which also prevents underwater object detection algorithms from fully leveraging their potential. In this article, we propose a self-supervised learning network for underwater target detection. Considering the deficiency that the conventional contrastive learning network pays more attention to the global information and ignores the local information, an auxiliary branch inspired by masked autoencoders is added to the baseline SimSiam network, which collaborates with the main branch to optimize the target network and help the target network learn the local information of the target feature map. A residual spatial cooperative attention module is proposed to be embedded within the proposed self-supervised learning network to obtain remote information through residual structure and construct spatial context features. The method of cooperative attention is used to enhance feature learning ability. Experiments are carried out on a reconstructed underwater target data set. Results show that compared with the baseline network, the method proposed in this article is more suitable for underwater environments and has better mean average precision.