A Multi-Scale Defect Detection for Steel Surface Based on Improved Faster R-CNN

The accuracy of defect detection is increased by the improved Faster R-CNN detection algorithm, which addresses the issue of an extended target scale span and complex characteristics causing missed detection in steel surface defect detection. The algorithm introduces path aggregation network (PANet), which uses bottom-up path aggregation to combine shallow details and deep semantic information to better capture features on various dimensions. It also adapts the reconstruction of the anchor box size information to filter out the optimal proposal box for more accurate localization. The backbone network of the original Faster R-CNN is replaced with an improved ResNet50, which has more powerful feature extraction skill and makes the model more flexible. On the NEU-DET dataset, the improved algorithm's detection performance is contrasted with that of other detection algorithms. According to the results of the investigation, the Faster R-CNN with additions has an average detection rate of 63.7 frames per second (FPS) and a mean average precision (mAP) of 80.2%, which is 8.7 percentage points higher than that of the Faster R-CNN. The conclusion is that the boosted Faster R-CNN can enhance both the precision and localization information of multi-scale faulty targets.