Steel Defect Detection Using Improved Multi-Head Attention and Dynamic Loss Re-Weighting in Cascade R-Cnn

This paper proposes a Cascade R-CNN steel defect detection method based on improved multi head attention mechanism and dynamic multi-stage loss reweighting, and validates it on the Northeastern University steel surface defect dataset. This method enhances the feature expression ability of the model for key defect areas by introducing an improved multi head attention mechanism. By combining dynamic multi-stage loss reweighting strategy, the robustness and detection accuracy of the model have been improved. According to the experimental results, the detection accuracy of the improved method has increased by 2.6 % compared to the original Cascade R-CNN, and the use of the improved method can effectively improve the efficiency of steel defect detection.