A defect detection approach based on Gabor wavelets is proposed. Acquired image is multi-channel filtered with a bank of 2-D real Gabor filters. Next

feature images are obtained by subjecting each filtered image to a nonlinear transformation and computing a measure of energy through smoothing filter. In the learning phase

statistical parameters are computed using learning sample which is free of defect and are used to supervise the thresholding of inspected image. Finally

threshold-fietered feature images are fused inter-scale and inter-orientation and binarized in order to decrease false alarms. Experiments demonstrated that this approach has good detection ability performance and needs less learning samples

which makes it suitable for many types of defect and textured material.