Fabric Defect Detection via Un-Decimated Wavelet Decomposition and Gumbel Distribution Model

Abstract

This paper investigates an unsupervised approach for fabric defect detection using un-decimated wavelet decomposition and simple statistical models. A novel data fusion scheme is proposed to merge the information from the different channels into a unique feature map in which potential defective regions will be highlighted distinctly. The distribution of the pixel values corresponding to the defect-free background texture in the feature map is modeled as per the Gumbel distribution model whose parameters are estimated by partitioning the feature map into a set of small patches. By calculating the log-likelihood value of each patch, a log-likelihood map (LLM) can be conveniently created, which provides a good cluster representation of the non-defective regions. A simple thresholding procedure then follows to discriminate between defective regions and homogeneous background in the LLM. The performance of the method has been extensively evaluated using a variety of real fabric samples, and the effectiveness of the proposed scheme has been verified by experimental results in comparison with other methods.