Lightweight Tunnel Defect Detection Algorithm Based on Knowledge Distillation

One of the greatest engineering feats in history is construction tunnels, and management tunnel safety depends heavily on detection defects. However, real-time, portability, accuracy issues with present defect technique still exist. The study improves traditional technology based knowledge distillation algorithm, depth pooling residual structure designed teacher network to enhance ability extract target features. Next, MobileNetv3 lightweight built into student reduce number volume model parameters. then trained terms both features outputs using a multidimensional approach. By processing radar photos, dataset created. experimental findings demonstrate that approach greatly increases efficiency: parameters decreased by 81.4%, from 16.03 MB 2.98 MB, while improved 2.5%, 83.4% 85.9%.