Object detection oriented underwater image enhancement based on gradient-guided generator and grouped spatial attention

Underwater absorption and scattering cause image quality distortion, which seriously affects the accuracy of object detection. Most existing methods do not aim to improve the accuracy of downstream machine vision tasks but to enhance the quality of human visual perception. In this paper, we propose an object detection oriented underwater image enhancement based on a gradient-guided generator and grouped spatial attention (GSA). Concretely, we propose a gradient-guided generator network to effectively represent the edges and details of the underwater images. Additionally, the GSA is designed to enhance feature representation by aggregating spatial context information and then improving the performance of the model in specific feature areas. We evaluate the proposed method on multiple underwater image datasets. Experimental results on the ChinaMM dataset show that our method achieves 98.3% on and 78.7% on AP. These results demonstrate that our method significantly improves object detection accuracy, and its overall performance is better than that of other methods.