Side-Scan Sonar Image Segmentation Based on Improved Deeplabv3 Plus

Sonar image segmentation is critical for improving the recognition and detection of underwater targets, particularly in complex environments. Although deep learning has shown promising results in image segmentation using convolutional neural networks (CNNs), obtaining an accurate boundary of the target remains a challenge as a result of the partial loss of details during down-sampling. In this paper, a novel improved model is proposed that incorporates an edge detection structure into the CNN architecture. The edge features are aggregated with class-specific semantic features extracted from deep networks through a channel attention mechanism, which further refines the fine segmentation of object boundaries. The experimental results demonstrate that our improved Deeplabv3+ algorithm achieves better performance in side-scan sonar image segmentation compared to other methods, highlighting its potential for improving the recognition and detection of underwater targets.