Underwater Object Detection With Mixed Attention Mechanism And Multi-Enhancement Strategy

Different from the real-world object detection tasks, underwater object detection has a more complex environment. The underwater environment is often faced with the challenges of the blur, scale variation, color shift, and texture distortion. Most generic object detection algorithms are difficult to maintain real-world robustness in the complex underwater scene. To address these problems, we propose a two-stage object detection method with an attention RPN block and an effective multi-enhancement strategy. For the case of blurred underwater images, our attention RPN improves the characteristic expression of key areas, which helps to gain better proposals in the region proposal stage. Besides, We also introduce a multi-enhancement strategy to improve the quality of the detected pictures. To boost the detection effect, our multi-enhancement strategy works on reducing the domain shift between underwater scene and real scene, especially the images with color shift and distortion. We do extensive experiments in the URPC datasets and MS COCO 2017. Finally, in underwater tasks, our method can gain 1.2 mAP higher than that of results in the baseline.