Abstract: Objectives Aiming at the high-intensity noise and large target obstacle structure characteristics faced by underwater sonar imaging process, as well as the stringent requirements of real-time underwater obstacle avoidance task for lightweight deployment of sensing algorithms and high inference efficiency, a semantic segmentation algorithm for sonar images with low computational overhead and short inference time characteristics is proposed to cope with the contradictory conflict between computational complexity of sensing algorithms and real-time response efficiency under the obstacle avoidance demand. Methods The perception algorithm significantly reduces the computational complexity by introducing the lightweight depth convolution operation, and at the same time, it is designed to effectively feel the field for the obstacle avoidance scenario, targeting to improve the accuracy of the large target segmentation, which fully meets the real-time perception task requirements based on the underwater obstacle avoidance scenarios of the forward-looking sonar. Results Through the training experiments with 6936 sonar images labeled by real collection, it is verified that the proposed perception algorithm reduces the reasoning time by 22.6% and improves the perception accuracy by 10.8% compared with the base model while the computational and parametric quantities are cut by 69% and 83%, respectively. In addition, by experimenting the obstacle avoidance framework based on the perception algorithm and model predictive control in the Gazebo simulation platform, the real-time and effectiveness of the perception algorithm in the obstacle avoidance process are verified. 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.