Criss-cross global interaction-based selective attention in YOLO for underwater object detection

With the development of computer vision, object detection has attracted wide attention and achieved exciting results in most situations. However, facing underwater environments, object detection’s performance degrades severely due to multiple ineluctable factors, including poor underwater imaging quality, underwater objects with protective colors, etc. These lead to strong interference of underwater backgrounds and the weak discriminability of underwater object features, which make underwater object detection become an extremely challenging task and cry out for reliable solutions. In order to reduce the underwater background interference and improve underwater object perception, we first propose the criss-cross global interaction strategy (CGIS). CGIS consists of two criss-cross structures, where feature decomposition and feature extraction are performed sequentially according to different criss-cross shapes in each structure. For information interaction, our strategy simultaneously avoids the destruction of direct information correspondence and the lack of global information interaction. According to different parameter allocation strategies, CGIS is further divided into standard criss-cross global interaction strategy (SCGIS) and efficient criss-cross global interaction strategy (ECGIS). We then design the criss-cross global interaction-based selective attention in different target dimensions. Our selective attention efficiently perceives global underwater information and rationally allocates precious computing resources to important underwater regions. We finally combine the designed selective attention with YOLO detectors, where attention modules are added to both ends of the feature fusion. The experimental results show that our work makes important progress in achieving efficient underwater object detection. Our selective attention shows good robustness in various YOLO detectors and exhibits ideal generalization in different detection tasks.