A weld defect segmentation framework based on visual self-supervised learning

When a ray image is selected as the target object, accurately segmenting the internal defects of a weld can yield the size and contour information of the defects and provide a basis for evaluating the welding quality. Currently, the mainstream defect segmentation strategies are image semantic segmentation methods based on standard supervised learning. As data-intensive paradigms, such methods are usually supervised by pixel-level dense labels, which not only require the labellers to have expertise in radiographic inspection and experience in defect assessment but also involve a very time-consuming labelling process. Despite the large amount of defect data accumulated in the current industry, the amount of available labelled data is small. To solve this problem, an automatic weld defect segmentation framework based on self-supervised learning theory is proposed. Specifically, a semiautomatic defect data extraction engine was constructed to form a large defect dataset that can be used for self-supervised pretraining, and an object-level masking strategy and a multihead decoding segmenter were designed. Experiments show that the proposed method can effectively utilise unlabelled data and improve the performance of the defect segmentation model. The defect data used in this paper can be obtained at https://github.com/longteng-coder/Wdsf-ssl.