Real-Time Segmentation of Side Scan Sonar Imagery for AUVs

Sonar imaging is considered the only feasible sensing solution for underwater investigations by autonomous underwater vehicles (AUVs) in zero-visibility water conditions. For AUVs equipped with side-scan sonar (SSS), real-time automatic segmentation of the sonar imagery can assist its navigation and path planning, thereby enhancing its autonomy. Therefore, the ability to perform pixel semantic segmentation in real time is critical when AUV performs tasks. However, due to the problem of speckle noise and intensity inhomogeneity in sonar images, robust sonar image segmentation methods are rare. Convolutional neural networks (CNNs) have been developed rapidly recently for its outstanding ability in image classification and strong robustness. Consequently, this paper uses a new CNN network architecture called ENet to segment the SSS image. In order to reduce the impact of class imbalance problems, a training strategy based on data augmentation and weighted loss function is adopted. Furthermore, to take into account the local dependencies of class labels, the results of ENet are used to initialize MRF to obtain the final segmentation maps. We have tested the proposed method on the SSS image dataset and reported a comparison with other image segmentation methods. The results show that the method can ensure the efficiency and accuracy of target recognition and provide target information for AUV, thus ensuring the high autonomy of AUV.