De-DETR: A More Accurate Real-Time Defect Detection Method

In industrial production, defect detection technology plays a very important role in improving product yield and increasing enterprise efficiency. At present, defect detection technology based on deep learning mainly focuses on convolutional neural networks. With the gradual application of the Transformer architecture represented by the DETR[1] model in the image field, there is a new direction for the update of defect detection technology. This paper proposes a defect detection method based on DETR. This method focuses on improving the encoding method of defect information, further improving the information fusion efficiency of deep features and shallow features, while maximizing the retention of deep semantic information, effectively improving the overall performance of the model. Experiments show that this method is superior to excellent real-time detection models such as YOLO and Rt-DETR[2] in terms of defect detection effects on the NEU-DET dataset.