Steel surface defect detection approach via fusing multiple hierarchical features
The proposed multi-level feature fusion network combines multiple hierarchical features into one feature.Based on multi-level features,a region of interest(ROI)is generated using a region suggestion network,and the final detection result is gen-erated by a detector composed of a classifier and a bounding box regressor.Establish a defect detection dataset HGSZ-DET for training and evaluation.Using 120 samples,mAP was achieved 75.1%on the baseline network ResNet18,and 84.6%on the base-line network of ResNet50.In addition,by using only 50 samples,this method can detect at a speed of 31 feet/second on a single GPU and achieve 95%of the above performance,with the potential for real-time detection.
万方数据
