Sonar image segmentation is an integral part in the development of image segmentation technology. It is also an important part of underwater target recognition and detection. In traditional methods，the algorithms based on supervised segmentation methods are often expensive，and are characterized with long test periods，poor real-time performance，and low running rates. In addition，due to objective factors such as poor imaging quality，low resolution，unclear edge conditions，and large workload of manual annotation of sonar images，it is difficult to establish large-scale datasets for supervised model training. Therefore，traditional segmentation methods are becoming less and less suitable for various requirements of current practical applications. In this paper，the unsupervised learning convolution neural network (UCNN) is introduced into the sonar image segmentation task. The segmentation model is then trained and tested on a single frame of sonar images. Finally，by inference，a sonar image is obtained after segmenting the shadow area and the target highlight area，and the segmented underwater target is obtain. Through the analysis of various indicators of the segmentation results of the experiment，it is proved that this method has better operation efficiency and segmentation accuracy，as well as higher real-time performance. It also has better comprehensive performance compared with traditional methods.
