Metal Surface Defect Detection Based on Transformer Merging Edge Information

Defects of metal surface can influence quality of industrial products. It can not only affect the appearance but also the physical quality of the product. Detecting defects of metal surface can efficiently reduce the impact of defects on product quality. Traditional detection methods require special equipment, which not only has low accuracy but also is expensive. Method based on vision has low cost and high efficiency. Because high detection precision and recall is required, method based on deep learning is used widely recently. Research on deep learning models is getting hotter and hotter. This paper designs a novel model--Transformer Merging Edge Information (TMEI), which concatenates edge information extracted by edge detector and input images. Deep abstract characteristics are extracted from these information-rich images. Transformer Encoder is used to utilize weighted blending of information across distinct regions. Lastly, feed forward net is used to classify defects. Experimental results show TMEI has an moderate advantage in precision, recall and F1-score.