Automatic surface inspection using wavelet reconstruction

In this paper, we present a multiresolution approach For the inspection of local defects embedded in homogeneous textured surfaces. The proposed method does not rely on the extraction of local textural features in a pixel-by-pixel basis. It is based on an efficient image restoration scheme using the wavelet transforms. By properly selecting the smooth subimage or the combination of detail subimages in different decomposition levels for backward wavelet transform, the reconstructed image will remove regular, repetitive texture patterns and enhance only local anomalies. Then, a simple thresholding can be used to discriminate between defective regions and homogeneous regions in the reconstructed image. This converts the difficult defect detection in complicated textured images into a simple binary thresholding in nontextured images. Effects of different wavelet bases, the number of multiresolution levels, the decomposed subimages used for reconstruction, and the change in image rotation on defect detection are thoroughly evaluated and discussed based on the experiment on a variety of real textures including machined surfaces, natural wood, textile fabrics, sandpaper and leather. (C) 2001 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.