Underwater Target Recognition via USNs

This chapter introduces several object recognition methods with different underwater information, including one-dimensional acoustic signals, sonar images and optical images. For one-dimensional signals, Siamese Network and few shot methods are adopted. For sonar images, Grad-CAM and style transfer methods are applied to improve the object localization and classification performances. For optical images, YOLOv4 network is employed to provide higher recognition accuracy. All these methods show their advantages over other methods, and tend to be an efficient and accurate method in their corresponding fields. Moreover, this chapter commences a type of object recognition method using information fusion of sonar and optical images, so as to make full use of detected information, and achieve better recognition performances. This method is based on late fusion, and Detection Transformer (DETR) is included in the fusion task. The result implies that fusion methods can join the benefits of both kind of images, and is conducive to recognition tasks.