Methods for Improving the Efficiency of Object Detection in Sonar Images

The presented work addresses the issues of improving efficiency in solving the problem of detecting and identifying objects using neural network methods on sonar images. It is shown that it is possible to apply methods developed for optical images — both classical and modern neural network ones — when implementing the process of secondary data processing for a sonar image. The methods of manual and automatic object detection in a sonar image are considered. The disadvantages of these methods are indicated. The issues related to the use of High-resolution neural networks for test and training sets of convolutional neural networks used for object detection and identification are discussed in detail. As part of a numerical experiment on real data, it is shown that the use of High-resolution neural networks as part of the detector allowed increasing the object detection efficiency by 10 percent and completeness by 10 percent. The most effective is the object detector on the sonar image, for which the High resolution BSRGAN method is used at the training stage of the Yolo3 neural network.