Objectives In view of the high-intensity noise and large-target obstacle structure characteristics faced in the underwater sonar imaging process, as well as the strict requirements for lightweight deployment and high inference efficiency of perception algorithms in real-time underwater obstacle avoidance tasks, a semantic segmentation algorithm for sonar image with low computational overhead and short inference time characteristics is proposed to deal with the contradiction between the computational complexity of the perception algorithm and the real-time response efficiency under the obstacle avoidance requirements.

Methods Based on the encoder-decoder network structure, this paper significantly reduces the computational complexity by introducing lightweight convolution operations, and at the same time introduces large separable kernel attention into the skip connections for obstacle avoidance scenarios. The 6936 sonar images collected and annotated in the real scene were trained and compared, and the obstacle avoidance strategy based on the perception algorithm was verified on the Gazebo simulation platform.

Results The modified algorithm specifically improves the segmentation accuracy of large targets. Compared with the benchmark model, the FLOP and parameters are reduced by 69% and 83%, respectively. At the same time, the inference time is reduced by 22.6%, and the accuracy is increased by 10.8%.In addition, simulation experiments verify the effectiveness of the perception algorithm in the obstacle avoidance process, and fully meet the needs of real-time perception tasks in underwater obstacle avoidance scenarios based on forward-looking sonar.

Conclusions The proposed perception algorithm based on sonar images can effectively solve the obstacle avoidance needs of unmanned underwater vehicle in airborne scenarios and has good engineering application prospects.