Statistical Approach to Unsupervised Defect Detection and Multiscale localization in Two-Texture Images

We present a novel statistical approach to unsupervised detection and localization of a chromatic defect in a uniformly textured background. The test images are probabilistically modeled using Gaussian mixture models, and consequently defect detection is posed as a model-order selection problem. The statistical model is estimated using a modified Expectation-Maximization algorithm that aids in faster convergence of the scheme. A test image is segmented only if a defective region/blob has been declared to be present, and this improves the efficiency of the entire scheme. This work places equal emphasis on defect localization; hence, an elaborate statistical multiscale analysis is performed to accurately localize the defect in the image. The underlying idea behind the multiscale approach is that segmented structures should be stable across a wide range of scales. The efficacy of the proposed approach is successfully demonstrated on a large dataset of stained fabric images. The overall detection rate of the system is found to be 92% with a specificity of 95%. All of these factors make the proposed approach attractive for implementation in online industrial applications.