The automated visual inspection of homogeneous flat surface products is a challenging task that needs fast and accurate algorithms for defect detection and classification in real time. Multi-directional and Multi-scale approaches, such as Gabor Filter Banks and Wavelets, have high computational cost in addition to their average performance in defect characterization. This paper presents a novel implementation of a neighborhood-preserving approach for the fast and accurate inspection of fine-structured industrial products using a new neighborhood-preserving cross-correlation feature vector. The fast and noise immune Probabilistic Neural Network (PNN) classifier has been found to be very suitable for defect detection in homogeneous non-patterned surfaces with acceptable slight variations, such as textile fabrics. A defect detection accuracy of 99.87% has been achieved with 99.29% recall/sensitivity and 99.91% specificity. The discriminant power shows how well the PNN classifier discriminates between normal and abnormal surfaces. The experimental results show that the proposed system outperforms the Gabor function-based techniques.
Defect detection in textile fabrics with optimal Gabor filter and BRDPSO algorithm
