To address the challenges posed by high-intensity noise and the structural characteristics of large obstacle targets in underwater sonar imaging, as well as the stringent requirements for lightweight deployment and high inference efficiency of perception algorithms in real-time underwater obstacle avoidance tasks, a semantic segmentation algorithm for sonar images with low computational cost and short inference time is proposed. The method aims to resolve the trade-off between the computational complexity of perception algorithms and the real-time response requirements in obstacle avoidance applications.

Based on an encoder-decoder network architecture, lightweight convolution operations were introduced to significantly reduce computational complexity. In addition, a large-kernel separable attention mechanism was incorporated into the skip connections to enhance feature fusion for obstacle avoidance scenarios. A dataset of 6936 sonar images collected and manually annotated from real environments was used for training and comparative experiments. Furthermore, the obstacle avoidance strategy based on the proposed perception algorithm was validated on the Gazebo simulation platform.

The improved algorithm specifically enhances the segmentation accuracy of large targets. Compared with the baseline model, the FLOP and the number of parameters are reduced by 69.2% and 83%, respectively. At the same time, the inference time is shortened by 22.6%, while perception accuracy improves by 10.8%. In addition, simulation experiments verify the effectiveness of the perception algorithm during the obstacle avoidance process, demonstrating that it fully satisfies the requirements of real-time perception tasks in underwater obstacle avoidance scenarios based on forward-looking sonar.

The proposed sonar-image-based perception algorithm can effectively meet the obstacle avoidance requirements of unmanned underwater vehicles in onboard operating scenarios and shows promising potential for engineering applications.