A Combined Weld Net-UNet Segmentation Model for Welded Seam Detection

Nowadays, Weld Seam (WS) detection is important in the field of intelligent manufacturing of welding as well as in the automated detection of defects. However, for the WS detection, point clouds are used as a major aspect; while the existing models used Combined UNet and WeldNet segmentation for WS detection, but individually UNet does not support tube sheet point clouds while individually WeldNet does not support regular point clouds. Therefore, in this paper, the integration of WeldNet-UNet segmentation model is proposed for the effective detection of WS. Firstly, WS dataset is preprocessed using sparse convolutional middle layers to generate 3D-feature maps from input images. After that, these feature maps are further processed with pre-trained WeldNet model followed by UNet encoder to extract voxel features and corresponding UNet decoder generated binary segmentation output as defective and non-defective WS. From the results, the proposed WeldNet-UNet model performed better compared to existing UNet and Zoom in on Target (ZIOT) in terms of accuracy (98.53 %), jaggard coefficient (96.77 %), Dice Similarity Coefficient (99.58 %) and recall (99.67%) respectively.