Recognition and classification of passive sonar targets using semi-supervised deep learning, attention gates, and jump connections.

Due to the critical role of sonar signals in identifying and tracking targets and objects in the water environment, the existence of classification systems for these targets is particularly important. There are various ways to do classification. One of these ways is using automatic systems that intelligently classify targets, which are less expensive and can perform the classification process with high accuracy.In this research, the main goal is to automatically classify targets in the water environment using passive sonar signals and through semi-supervised deep neural networks. In this research, several types of neural network architectures have been presented, which, by using new mechanisms such as the use of attention gates and jump connections, have high classification accuracy compared to existing methods and have acceptable performance in performing classification. By using only 25% of the labeled data, the proposed basic network has provided a 10.59% improvement compared to the accuracy stated in the article of the examined data, with an accuracy of 70.17% on the test data. Also, the other proposed network called Attention-Resnet-2 has exceeded by obtaining 71.92% accuracy and has achieved this improvement by 12.34%.