Inspection of Steel Strip Surface Defects Based on Visual Attention Mechanism

Abstract

Aiming at the low contrast and small defects on the surface of steel strip, a defect detection method based on human visual attention mechanism is proposed. On the one hand, the improved Itti model is used to obtain the bottom-up saliency map, that is, the Gaussian pyramids of intensity, orientations and edges are constructed by multi-scale filtering of the image, and the “center-surround differences” operation and normalization are used to form the intensity, orientations and edges conspicuity maps. Then, each kind of conspicuity maps is combined into bottom-up saliency maps by weighting method. On the other hand, the multi-channel and high-level features are extracted by convolutional neural network (CNN), and the high-level saliency map is formed by fusion and up sampling. Next, the comprehensive saliency maps are generated by weighting bottom-up saliency maps and high-level saliency maps, so as to realize the comprehensive fusion of image features. The results show that this method has good effect on low contrast and small defects on the surface of steel strip.