High Accuracy Vision-Laser Post-Welding Defect Detection and Segmentation Method Based on Convolutional Neural Networks

Recently, the advent of deep learning technology has led to the exploration and application of Convolutional Neural Networks (CNN) based welding defect detection methods by numerous researchers in several industrial settings. However, most CNN-based methods are single-task networks that can only classify the type of welding defect, lacking the capacity to provide additional information such as the shape of the defect. This limitation hinders their application in more diverse scenarios. This paper proposes a high-accuracy, CNN-based method for post-welding defect detection and segmentation using the Vision-laser system. The proposed method, which comprises a multitask CNN, can simultaneously detect, classify, and segment the area of the welding defect. This approach allows for the acquisition of more comprehensive information about the welding defect compared to existing single-task methods. A Vision-laser system is assembled and calibrated for data collection and testing. The proposed network is trained and evaluated on this dataset. The results demonstrate that the method achieved a mean accuracy of 99.58%, a mean precision of 99.43%, and a recall of 97.05% on the welding defect detection task. Additionally, it achieved a mean Intersection over Union (IoU) of 92.58% on the welding defect segmentation task. The proposed method exhibits high accuracy in both defect detection and segmentation tasks.