Fast defect detection in homogeneous flat surface products

This paper introduces a novel hybrid approach for both defect detection and localization in homogeneous flat surface products. Real time defect detection in industrial products is a challenging problem. Fast production speeds and the variable nature of production defects complicate the process of automating the defect detection task. Speeding up the detection process is achieved in this paper by implementing a hybrid approach that is based on the statistical decision theory, multi-scale and multi-directional analysis and a neural network implementation of the optimal Bayesian classifier. The coefficient of variation is first used as a homogeneity measure for approximate defect localization. Second, features are extracted from the log Gabor filter bank response to accurately localize and detect the defect while reducing the complexity of Gabor based inspection approaches. A probabilistic neural network (PNN) is used for fast defect classification based on the maximum posterior probability of the Log-Gabor based statistical features. Experimental results show a major performance enhancement over existing defect detection approaches.