Steel Surface Defect Detection Using a New Haar–Weibull-Variance Model in Unsupervised Manner

Automatic defect detection on the steel surface is a challenging task in computer vision, owing to miscellaneous patterns of the defects, low contrast between the defect and background, the existence of pseudo defects, and so on. In this paper, a new Haar-Weibull-variance (HWV) model is proposed for steel surface defect detection in an unsupervised manner. First, an anisotropic diffusion model is utilized to eliminate the influence of pseudodefects. Second, a new HWV model is established to characterize the texture distribution of each local patch in the image. The proposed model can project the texture distribution of each patch into the low-dimensional space with only two parameters. The parameter distribution of the whole image can also be unified into the form of linear radiation in an Euclidean space. The reliable background can be extracted via the formation of parameter distribution, by which the model parameter can be optimized further. Finally, the adaptive threshold can be determined to distinguish the defect from the background. Experimental results show that the proposed method can detect an arbitrary type of defect on the homogeneously textured surface and achieve an average detection rate of 96.2% on the data set, which outperforms the previous methods.