Research on hot rolled steel strip defects image processing method based on multi-scale detection

Abstract

Hot rolled steel strip is usually detected by image processing for defects, but it is often affected by light and conventional image processing methods cannot effectively detect defects with small areas. In this paper, an image processing method is proposed to overcome the effect of fill light for the detection of steel strip surface damage by convolutional neural network multiscale detection method.In the target detection part, K-means++ clustering was performed on the anchor size, while the S-SPP module for multi-scale detection was introduced to further improve the detection of small-area damage. After training the model with the NEU-CLS dataset, the accuracy reached 94.6% and the detection speed was 27.3 ms. Through comparison experiments, the experimental results show that the proposed method of image processing can detect steel strip damage more accurately and faster, which provides a feasible method for practical application in factory scenes.