Defect Detection in Random Colour TexturesXianghua Xie, Defect Detection in Random Colour Textures. PhD thesis. University of Bristol. April 2006. No electronic version available.
Automated surface inspection for quality control has largely employed graylevel image processing techniques, for example in textile and wafer inspection. There are rising demands in the quality control industry for colour analysis to fullfil its vital role in visual inspection, e.g. in ceramic tile manufacturing. This thesis is concerned with developing texture analysis techniques in application to the detection of abnormalities in colour texture surfaces, in particular ceramic tile surfaces on which patterns are regularly of a random nature. These abnormalities can be divided into two categories: colour tonality defects and textural abnormalities.
Colour tonality refers to global chromatic appearance of a surface. Its variation from surface to surface may be understated, but becomes significant once the surfaces are placed together. In the first part of this thesis, a multidimensional histogramming method is presented to detect subtle tonality changes by incorporating local chromatic information into global chromatic characteristics. PCA is used to reveal the nonlinear noise interference introduced by the imaging system. Vector directional processing and reference eigenspace feature selection are proposed to obtain salient colour tonality representation. The method is evaluated on a dataset with groundtruth, and compared against an existing state-of-the-art method.
Textural quality inspection involves the detection and localisation of various chromatic and textural imperfections. This thesis suggests that although some textures have a random appearance, there are textural primitives that govern the global appearance. A novel two-layer generative model is proposed to represent an image or a family of images. In this model, random (or regular) texture images in the first layer are assumed to be generated from a collection of texture exemplars, or texems, in the second layer. A bottom-up texem generation process is proposed based on pixel neighbourhoods. This local contextual analysis using texems is applied to a large set of graylevel ceramic tile images, in which graylevel analysis is sufficient to detect textural abnormalities. Then, different schemes are explored to extend graylevel texems to colour images. This results in two different formulations and inference procedures with different computational complexity.
Spatial detection and localisation accuracy of the proposed methods is measured and compared using texture collage images. The texem model is also compared against the multiscale, multidirectional Gabor filter for defect detection. Both the colour tonality inspection and textural defect detection methods are implemented in novelty detection schemes to cope with the variety and unpredictable nature of defective samples.
An application of the colour texem model to medical image analysis, involving the detection of abnormalities in tympanic membrane images, is briefly discussed. Finally, a further extension of the texem model to perform segmentation on inkjet printed ceramic tile surfaces is presented.