Side scan sonar segmentation using deep convolutional neural network

Side scan sonar (SSS) is a vital sensor for autonomous underwater vehicles (AUVs) to do ocean survey. Many methods have been proposed to carry out SSS image segmentation, among which machine learning algorithms provide outstanding performance. Machine learning algorithms like support vector machine (SVM) and convolutional neural networks (CNN) are the most used. When SVM is used to do pixel-level classification or segmentation, image features should be chosen based on researchers' experience. So the result may be not the optimal one. On the other hand, deep CNN can extract features of images automatically through its special architecture and can capture highly nonlinear mappings between inputs and outputs. Therefore, in this paper, we propose to use deep CNN to segment images of SSS into three parts: highlight areas with objects, regions of shadow and sea-bottom reverberation areas. The CNN used in this paper is derived from fully convolutional networks (FCN) which is trained by manually segmented sonar data, takes patches of the same size as inputs while training and outputs the segmentation results of images directly. It is an end-to-end architecture which can output segmentation results of side scan sonar images with any size after training. Then we post-process the results of CNN by using Markov random fields (MRF). Experimental results demonstrate that segmentation method based on CNN and MRF is applicable in SSS image segmentation.