Defect detection of hot rolled steels with a new object detection framework called classification priority network

Abstract

The defect detection is extremely crucial to analyze the surface quality of steel products, it not only affects the subsequent production, but also affects the corrosion and wear resistance of the end products. However, the low classification and detection rates of the conventional algorithms cannot satisfy the demands of some steel production lines, such as hot rolled plates. To improve the accuracy of defect inspection, we present a new object detection framework, classification priority network (CPN), and a new classification network, multi-group convolutional neural network (MG-CNN), to inspect the defects of steel surface. In CPN, the image is first classified by MG-CNN, which trains different groups of convolution kernels separately to extract the feature map groups of different types defects. Then, according to the classification result, the feature map groups that may contain defects is separately input into another neural network, which is build based on yolo, to regress the bounding boxes of the corresponding defects. Extensive experiments are carried out on samples from different steel production lines indicate that the proposed method had yielded best performances compared with prior methods used in defect inspection. The average detection rates of surface defects in hot rolled plates and strips are higher than 94%, and the classification rates of in above production lines are above 96%.