Hybrid Modeling Based Semantic Segmentation of Forward-Looking Sonar Images

Semantic segmentation of forward-looking sonar (FLS) images plays a key role in the perception and interaction of autonomous underwater vehicles with the surrounding environment. Due to the strong noise and blurred object edges in sonar images, there is a high demand for the model's feature extraction and anti-interference ability. Currently, most methods are based on convolutional neural networks (CNNs), which are sensitive to local noise, and have a heavy computational burden, making them difficult to meet real-time requirements. This article re-examines CNNs and vision transformers, proposing a hybrid modeling-based network called HMSeg that combines both convolution modeling and attention modeling approaches for sonar image segmentation. In addition, a dynamic attention gate module is proposed to dynamically enhance feature maps with high-level features and eliminate interference. Furthermore, we propose a composite loss function to guide the model in extracting pure features and accurate semantic information. We present a new FLS image data set and conducted a series of experiments on a marine debris data set and a UATD-Seg data set. The results demonstrate that our proposed HMSeg achieves the best performance, proving its robustness and efficiency in different environments.