Identification of surface defects in textured materials using wavelet packets

This paper investigates a new approach for the detection of surface defects, in textured materials, using wavelet packets. Every inspection image is decomposed with a family of real orthonormal wavelet bases. The wavelet packet coefficients from a set of dominant frequency channels containing significant information are used for the characterization of textured images. A fixed number of shift invariant measures from the wavelet packet coefficients are computed. The magnitude and position of these shift invariant measures in a quadtree representation forms the feature set for a two-layer neural network classifier. The neural net classifier classifies these feature vectors into either of defect or defect-free classes. The experimental results suggest that this proposed scheme can successfully identify the defects, and can be used for automated visual inspection.