Segmentation of Side Scan Sonar Images on AUV

Side scan sonar is an active sonar. It uses seafloor backscattering to obtain the topographic information of the seabed, and constructs seabed topographic image information, which is the basis of seabed imaging. For AUV equipped with side scan sonar, the interpretation of sonar images is beneficial to enhance their autonomy. Based on efficient and accurate segmentation algorithm, the efficiency and accuracy of target recognition can be ensured. Therefore, how to automatically extract and understand marine information in acoustic images becomes the current research hotspot. Related work such as classification of marine targets, detection of underwater targets, etc. Real-time image acquired by the side scan sonar carried by AUV. In order to better identify the seabed information, we combine data augmentation, super-resolution technology, deep learning, and Markov random field(MRF). First, perform data preprocessing, getting marine data sets is very difficult and expensive. For small data sets are not sufficient to support deep network, and in order to prevent overfitting, we have performed data augmentation. Second, because the data collected by the sonar contains more noise, in order to achieve better image recognition, we introduced a super-resolution technology to process the low-resolution image into a high resolution image. Third, we perform image segmentation for the processed data. During this process, the resolution of the data collected by the side scan sonar is very low. In order to avoid the images as the network deepens, receptive field is getting smaller and smaller, so that more noise is added, we have adopted the DeepLab network framework. Finally, in order to better reflect the generalization ability of the network, so that the network has better performance, we take into account the label information of the nearby image, the Markov random field(MRF) was introduced at the back end. Experiments show that this method has better results.