A light-weight object detection method based on knowledge distillation and model pruning for seam tracking system

In the seam tracking process based on laser vision, the camera continuously collects weld seam images and locates weld feature points by utilizing image processing algorithms. It is critical to extract the feature points of the weld seam accurately and in real-time from noise interference. Deep learning-based weld seam recognition methods have high accuracy and strong robustness, but it is hard to satisfy the real-time requirements when they are deployed on devices with low computing power. To solve this problem, a lightweight object detection network based on Single Shot MultiBox Detector is proposed. Then the proposed method is in comparison with the mainstream deep learning-based algorithms, and the welding experiments are executed. The experimental results show that the average localization error is within +/- 0.2 mm, and the image processing speed reaches about 49 FPS on the CPU, demonstrating that the proposed network could satisfy the requirements of accuracy and real-time.