Weld Detection and Tracking Algorithm for Inspection Robot Based on Deep Learning

Due to the particularity and security of the special equipment, its welds should be inspected regularly. Compared to the manually operation mode, the inspection robot are able to improve the inspection efficiency. However, there are still technical problems in the accurate detection and tracking of the weld for the inspection robot. To the illustration, we develop an efficient algorithm based on the deep learning to the weld detection and tracking. In this algorithm, the improved YOLOv5n instance is used to segment the network for the weld detection. For the Backbone, the MixConv convolution module is introduced to replace the Conv module of the original network, which provides a larger receptive field and richer semantic information for the network. To improve the computational speed, the Ghost-C3 structure is introduced to replace the C3 structure in the Neck of the original network. In addition, the least square method is used to segment the weld to extract the weld trajectory with high accuracy. The experimental results of our algorithm show that, the size of the improved segmentation model is only 3Mb, and the weight is 26.8% smaller than the original model. Moreover, the precision of the improved model reaches to 99.0%, which is 1.6% higher than the original model.