A Novel Modeling of Random Textures Using Fourier Transform for Defect Detection

Abstract

In this paper we are concerned with the problem of detecting defects on random texture surfaces. We propose a novel method for the addressed problem. Due to the nature of random textures, characterizing normal patterns from defects is difficult. In this method we use the approach so called Phase Only Transform from Fourier transform family to extract frequency features from the texture patches in training and test stages. In training stage we use the extracted features from the training image patches to learn the probability density function of patches in the feature space. The training is performed on non-defective training sample using Gaussian mixture model. In the test stage, we divide the test image into small patches and from each patch we extract frequency features similar to training images. We use weighted normalized Euclidean distance measure derived from the model parameters to set a proper threshold. In order to obtain a defect map, distance of feature vectors extracted from image under inspection at each pixel position is calculated against our learned model and compare with threshold. The result of experiments for detecting defects on random texture tiles is very promising.