Algorithms for vision-based defect detection on textured surfaces

Many industrial product surfaces can be viewed as specular surfaces. As the key part of this research, defect detection on smooth, specular surfaces is investigated. Using a specular plus diffuse reflectance model as the rational basis, a numerical investigation of the appearance of the topographical defects on smooth, specular surfaces is conducted. Image captured at diffuse angles was proposed as an effective means to detect topographical defects in smooth, specular coatings reliably. Furthermore, numerical investigations confirm that diffuse angle images yield a 3D characterization of surface defects, efficiently, from a single image.;Many other industrial product surfaces can be characterized as textures. Texture analysis plays a key role in methods of detecting defects on textured materials. This research investigates various texture defect detection methods in two categories: texture feature extraction and classification. The most pervasive algorithms in both categories are reviewed.;To thoroughly measure and to compare the performance of these algorithms in the domain of texture defect detection, a framework is constructed. It contains test images, quantitative measures and implementations of the algorithms. Extensive comparative studies are conducted using this framework by measuring the performance of algorithms. One-Against-All and cross validation strategies are applied to achieve more reliable and texture independent results.;A novel method for texture defect detection, Selective Gabor wavelet features, is proposed. As opposed to conventional Gabor wavelets which use a full set of Gabor filter banks, a filter selection scheme is applied to select a subset of the filters which have the most discriminating power. It results in a reduction of more than 70% of the computational cost on feature extraction.;A rotation and scale invariant texture defect detection method is proposed based on local measures of micro-edge pixel densities. Radon transform is utilized to achieve rotation and scale invariance by estimating the primary direction and scale factor of the texture. Experiment results show that the estimation errors for both primary direction and scale factors are well below 0.5%. Multidirectional micro-edge density features are proposed to represent the texture. The results on complex real fabrics show the efficiency and robustness of the proposed method.;KEYWORDS: Texture Defect Detection, Texture Feature Extraction, Classification, Rotation and Scale Invariance, Physics-based Image Synthesis...