Unbalanced Underwater Sonar Image Target Detection Based on Weak Contrast Feature Enhancement

Abstract

Underwater target identification, as a core technology for underwater unmanned detection, plays a crucial role in both military and civilian applications. As the only signal that can be transmitted over long distances underwater, acoustic waves are widely used in current underwater target identification. Sonar image, as a graphical representation of acoustic waves, is also gradually be-coming a current research hotspot. The existing sonar im-age target detection framework is mainly based on optical image target detection model, which is difficult to solve the following two problems, first, the high similarity between target foreground and background in sonar image, the feature extraction process is very easy to lose the target feature information, especially the small target feature information, and second, the problem of model overfitting caused by small samples of sonar image data set and imbalance of sample data for each category given. In order to solve the above problems, this paper proposes an unbalanced small target underwater sonar image target detection model based on YOLOv5 model with weak contrast feature enhancement network, firstly, by adaptively enhancing large, medium and small target features in the feature extraction and feature fusion stages to improve the accuracy of the model in the target regression detection stage, and then by improving the loss function to make the model associate the originally mutually The model is then improved by improving the loss function so that the model can be associated with the localization and classification tasks that are originally independent of each other, while reducing the impact of the model due to the imbalance between positive and negative samples. The experimental results show that the detection performance of the proposed improved model reaches 96.9% on the existing public dataset UATD, which is a significant improvement in recognition accuracy compared to the advanced 2-baseline model.