A segmentation network based on residual blocks and multi-channel images

As an indispensable key process in the industrial chain, the demand of welding could continue to maintain stable growth. However, various defects could inevitably occur in the welding process, and undetected defects could directly affect the bearing capacity and service life of the welding device. Therefore, it is necessary to develop intelligent detection algorithms for non-destructive quality inspection of the welding products. In this paper, we convert the localization problem and category classification problem of weld defects into a pixel classification problem, and propose a u-shaped segmentation network based on residual blocks and the concatenated multi-channel images, called CatResUNet. First, we concatenate the original weld image and the reversed image to enrich the input data, and obtain a doubled-channel image. We then use residual blocks as the backbone of the encoder of the u-shaped network to extract the weld defect features from the obtained doubled-channel image, and use the skip connection to fuse the underlying features from the encoder and the high-level features of the decoder to ensure the accuracy and detail retention ability of defect segmentation. Finally, we use the convolutional layers instead of the fully connected layers to perform pixel-level classification. Experimental results demonstrate that appending the reversed image to the original image to expand the network’s input channels could enrich the information that was fed into the network, hence effectively reducing the rate of missed and false detections, it is a crucial aspect for weld defect detection. Compared with CatUNet which is a standard u-shaped network based on doubled-channel images, the introduction of the residual blocks can help our CatResUNet network to achieve better or basically equivalent evaluation indicators. Furthermore, compared with the other three approaches, the higher dice coefficient and IOU values achieved by our CatResUNet network indicate its enhanced sensitivity to small targets while maintaining excellent performance for large target segmentation tasks. This ensures that while preserving the model’s ability to segment large defects accurately, it also improves its sensitivity to small defects.