A Study Of Cluster-based Sidescan Sonar Image Segmentation With Unsupervised Feature Learning

Sidescan sonar detection devices are commonly used in marine activities, which present information by images. Sidescan sonar images are in low resolution and contain a lot of noises,and also the intensity range of sidescan sonar images is relatively narrow. So we can see that it is difficult to divide sidescan sonar images into targets and background. After analyzing the current mainstream image segmentation technologies and development trend,this paper makes a research on the general purpose clustering algorithms and the feature extracting methods based on unsupervised learning.First of all,we analyzed the sources of the noises in sonar images,and discussed about the sonar noise classification and noise models. Some state of art methods of denoising in optical images are used in these sidescan sonar images as a trial. The experimental results show that the method is effective for the sonar image in different scenes. On this basis, further analysis of the effect of denoising algorithms are discussed,and some possible methods to improve the results are introduced.Then according to the characteristics of side scan sonar image, We selected and introduced two kinds of texture features which are important and widely used at present, the Local Binary Pattern and the Haar Like features. We also tried to use the sparse auto-encoder,a deep learning algorithm,to learn the features of sidescan sonar images. Through the comparative analysis with the texture features, we find the features that obtained from learning algorithm has obvious advantages.Then this paper describes the clustering algorithms such as the K-means,fuzzy clustering, hierarchical clustering and spectral clustering in detail. Experiments were deployed on these algorithms with the features learned previously. The results were analyzed and compared with these algorithms which use the intensity of gray images instead of the features extracted by the learning algorithm.Finally, we try to use parallel computing methods, such as OpenMP and CUD A, to speed up the K-Means algorithm. Experiments show that the parallel algorithm has achieved a good effect of segmentation. It indicates that it is possible to deploy the algorithms in the real-time scenarios.