Automatic Target Recognition in Forward-Looking Sonar Images using Transfer Learning

Automatic target recognition (ATR) in the sonar image is usually divided into two steps: The localization part aims to locate the region which most likely contains the targets. The classification part uses the information from the located region to determine the targets' class. Classical methods complete the ATR task with two steps by different techniques. However, some papers have shown that CNN can realize the ATR task in one-step given sufficient training data. In this paper, we try to use a one-step CNN to automatically recognize (locate and classify) target in sonar images. For the problem of insufficient sonar image data which easily causes over-fitting, we try to find the common ground of CNN in extracting features from the optical image and sonar image. We use style-transfer method to help CNNs extract shape features from optical image data, thus the pre-trained CNN model can be used to improve the CNNs' ability for ATR tasks in the sonar image.