Unsupervised learning model for side-scan sonar image segmentation
In view of the problem that common segmentation methods only rely on gray attribute information,it is difficult to achieve accurate segmentation of complex objects in side-scan sonar images,while deep learning methods significantly rely on the current situation of samples,a new SSS image segmentation algorithm based on unsupervised neural network model was proposed in this paper.Firstly,feature extraction and feature fusion were carried out by multi-scale convolution.Then normalization was carried out to avoid under segmentation.The loss function considered the constraints of feature similarity and spatial continuity to reduce the influence of noise on segmentation and solve the discontinuity problem of segmentation results.According to the test set segmentation,the proposed method achieved good results in the measured data.The intersection over union in the simulated data was 94.3％,and the false positive rate was only 0.08％.Both indexes were significantly superior to other methods,which could effectively meet the needs of side scan segmentation,and is of reference for navigation safety,ocean engineering,underwater archaeology and other fields.
万方数据
维普
