Welding defects recognition based on DCP-MobileViT network

Welding defects recognition based on machine vision provides a reliable basis for robot arc welding. However, due to such factors as severe noise interference of welding images, slight morphological differences of same defect, and restricted computational capacity of on-site devices, it has been a challenge to learn distinguishable characteristics from various weld seam defects as well as to improve defects recognition accuracy and model generalizability, with a lightweight network. Consequently, we propose a welding defects recognition approach by designing a novel DCP-MobileViT network to address this challenge. First, the denoised image and its corresponding transmission map are obtained by Dark Channel Prior (DCP) algorithm and served as two inputs of the proposed network. Then, a dual-branch network is designed to adaptively extract and merge feature information of two input images through the convolution and transformer mechanism. Finally, the proposed DCP-MobileViT model is tested and compared with three other models using datasets from different welding scenarios. The results indicate that the DCP-MobileViT model achieves superior welding defects recognition accuracy compared to the other models, demonstrating its excellent generalizability in different welding scenarios.