Application of Improved YOLO V3 Algorithm for Target Detection in Echo Image of Sonar under Reverb

Abstract

Abstract aiming at the problems of low detection rate and high false alarm rate of small and medium targets in images, a YOLO V3 improvement method is proposed and applied to the detection of small targets. Because the small target occupies less pixels and the feature is not obvious, it is proposed to up-sample the 8 times downsampling feature map of the original network output, splice the 2 times upsampling feature map with the second residual block output feature map, and establish the feature fusion target detection layer with 4 times downsampling output. In order to gain More small target feature information is added to the second residual block DarkNe-t53 the YOLO V3 network structure. k-means clustering algorithm is used to cluster the number of target candidate boxes and the aspect ratio dimension. The results show that the improved YOLO V3 algorithm can effectively detect small targets and improve the recall rate and detection rate of small targets.