Side-Scan Sonar Image Segmentation Based on Multi-Channel Fusion Convolution Neural Networks

Side-scan sonar is an important application in the field of ocean exploration. Accurate segmentation of target regions in side-scan sonar images is a challenging issue due to the low-resolution and strong noise interference. To accurately and faster segment the different categories target in sonar image, a novel convolutional neural networks (CNNs) model is proposed in this study. Firstly, the deep separable residual module is used for target regions multi-scale feature extraction and suppression noise feature information interference, and the multi-channel feature fusion method is used to enhance feature information transfer of convolution layers. Secondly, the adaptive supervised function is used for pixel-wise classification of different categories targets. Finally, to improve model generalization ability and robustness, the adaptive transfer learning method is introduced in the model training process. We have performed extensive experiments on side-scan sonar image with different targets and scales. The experimental results show that the detection accuracy of the proposed method reaches 95.73%, which is outperforms other state-of-the-art methods on the side-scan sonar image segmentation tasks. Moreover, the method has fewer computational parameters, facilitating future deployment it to underwater mobile detection devices.