A new method-improved yolov5 for detection of the marine

Abstract

Due to the particularity of underwater environment, such as uneven distribution of light, interference caused by various underwater noises to imaging, and weak target features caused by the protective colors of underwater organisms themselves, the existing target detection technology is not ideal in underwater application scenarios. In this paper, the photography function of underwater robot was used to take photos of four objects, namely sea urchins, scallops, sea cucumbers and starfish, and the parts with obvious texture characteristics were selected to make a dataset URFD with 8,000 images. In order to ensure the adaptive ability of the detection algorithm to application scenarios, we selected the yolov5 algorithm with the best comprehensive performance by testing the performance comparison of various existing networks, and optimized and improved it. In order to reduce the redundant computation of convolution operations, the fasterNet structure is introduced. By optimizing the loss function of the original network structure and reducing the coupling degree of the original detection head, yolov5 dropped 2 points from floating point operation and improved the map by 2 points on the basis of the original performance. And the training time cost is reduced by 22%. The name of the improved algorithm is yolovsfd.