Underwater target detection algorithm based on multi-scale feature fusion

Abstract

Underwater target detection technology is an important way to realize the biological monitoring of the marine ranching. Nevertheless, since light absorption of the water and particles in the water will have a scattering effect on the light, the majority of the captured underwater optical images have low contrast and color bias, which blurs the local details of the measured biological targets. In addition, due to the mutual occlusion of underwater organisms, which will lead to a higher rate of missed targets detection. To address the above problems, we propose a new target detection algorithm for complex underwater environments based on the YOLOv7 (You Only Look Once Version 7). Firstly, in order to enhance the underwater image quality, while improving the color bias and reducing the image noise. By comparing different image enhancement methods, RGHS (Relative Global Histogram Stretching) is used to enhance the quality of dataset. Secondly, in order to realize the network's perception of biological detail features, the MIS (Multi-Gradient Interaction Structure) is proposed. The structural innovation aims to mitigate the challenges posed by target proximity and overlap. Finally, after the backbone feature extraction, we propose the multiscale feature fusion structure. It facilitates the cross-channel communication of semantic information, thereby enhancing the network's capability to express features. The experimental results indicate that our proposed algorithm has achieved a detection accuracy improvement of 1.35 % compared to the original YOLOv7 (You Only Look Once Version 7). Simultaneously, it has reduced instances of missed detections and misjudgments of underwater targets, surpassing common object detection algorithms.