Sonar-based underwater object detection utilizing highly reliable training data simulation method

Underwater object detection in sonar images requires a large number of images of target objects. For easy and accurate object detection, this paper propose a method using convolutional neural network (CNN) training with a sonar image simulator. Instead of simulating very realistic sonar images which is computationally complex, the training images are simulated by calculating only simple semantic information and adding randomized degradation effects to the simulated images. Images with randomized noise make CNN robust to noise. As a result, CNN can detect the target at sea, without training dataset consisting of real underwater sonar images captured at a field which are difficult to obtain.2