Yüzey Hata Tespiti için Sonuç Ağırlıklandırma Tabanlı Resnet Öznitelik Piramit Ağ Mimarisi

Abstract

Theory and Methods: In this study, Result Weighting-based Resnet Feature Pyramid Network (SA-ROPA) architecture was developed for surface defect detection. The proposed model, pretrained Resnet50 network architecture based on a transfer learning approach was used to deal with the problem of insufficient samples. In addition, a new approach called Result Weighting (SA), instead of directly combining features from different levels of the FPN architecture, was used. In this approach, four different defect detection results were obtained simultaneously using different levels of feature from the Resnet50 network. Then, the different detection results were combined in the depth dimension and the point convolution process was applied to obtain the final result prediction.Results: MT, MVTec-Texture and AITEX surface defect detection datasets were used to test the proposed SA-ROPA architecture. In experimental studies, the mIoU value obtained for the MT, MVTec-Texture and AITEX datasets using the proposed model was calculated as 79.92%, 76.37%, and 82.72%, respectively. As a result, the proposed model was achieved higher success than the state-of-the-art methods and previous studies. Conclusion:In the experimental results, it was observed that the surface defects were detected with high performance using the proposed SA-ROPA model. These results showed that the proposed model can be used for automatic surface defect detection in manufacturing systems. In addition, it was envisaged that this study in which extensive experimental studies were carried out, will lead to further work in the field of automatic surface defect detection.