Segmentation of Defects in Images of Steel Sheets Using Deep Learning

One of the important tasks of computer vision is non-destructive testing. This technology allows you to find faults and deviations from the norm by photo and video images. One of the applications of non-destructive testing technology is the analysis of metal products for defects. For this, computer vision algorithms such as convolutional neural networks, transformers, and others can be used. The paper considers the problem of defect segmentation in images. Four types of defects are identified for images of steel. Severstal data was relabeled using the Roboflow service. Segmentation models were trained and their comparative analysis was carried out. The results of the study showed that the transformer architecture models, namely SegFormer, provide the highest accuracy. In particular, the gain in terms of the Dice coefficient metric was about 1.5% when segmenting the test sample. However, the use of such an architecture leads to a slowdown in processing speed. It is shown that with the use of segmentation models it is possible to automate some of the tasks of monitoring steel products.