Reducing False alarm rate of Sonar image Detection via Self-training

Sonar image target detection is a key technology for underwater unmanned systems, but it faces the challenge of high false alarm rates, which limits the effectiveness and efficiency of underwater unmanned systems in performing tasks. Therefore, this paper proposes a method to improve the false alarm rate by self-training a deep learning detector on sonar images. This method automatically generates proxy classification tasks based on the sonar image target detection dataset. Through pre-training with these proxy classification tasks, the deep learning detector's ability to learn target and background features is enhanced, thereby improving its ability to distinguish between targets and backgrounds and reducing the false alarm rate. Experimental results show that, at the same or even higher recall rates, this method can reduce the false alarm rate by 3.91% compared to traditional transfer learning methods. This method achieves improved generalization performance of the deep learning detector without using external datasets, providing an efficient self-training approach for underwater target detection scenarios with small sample sizes.