Color Cast Detection Based On Feature Extraction And Color Segmentation

Color cast detection plays a decisive role in the quality detection of ceramic tiles, which is reflected in the fact that the degree of color cast is an important index to measure the quality of ceramic tiles. At present, domestic ceramic tile quality detection still relies on manual work mostly, which is time-consuming and laborious. However, fatigue, distraction, low discrimination of human eye and other factors leads to false judgment easily. As a result, detection errors and omissions happens, which degrades production efficiency and quality. Therefore, the importance of research on intelligent detection system is self-evident. This thesis stems from the project of science and technology development plans of Jiangsu Province——《Online intelligent ceramic tile quality detection and classification system based on human vision simulation)), in which several feature extraction and color segmentation algorithms are proposed or optimized. Meanwhile, these algorithms are applied in a number of national and provincial projects undertook by the laboratory, such as the National Natural Science Foundation (61105015), Jiangsu Natural Science Foundation (BK2010366), Support Program by Jiangsu Provincial Science and Technology Bureau (BE2011747), Project of Jiangsu Provincial Environment Protection Bureau (2009017).The thesis describes overseas and domestic research actualities of machine vision technology, elaborates the connotation and status of feature extraction and color segmentation algorithms in the implementation process of color cast detection, analyzes the existing problems and difficulties encountered in key algorithms, and then puts forward or improves relevant algorithms. The main research contents include Canny-Hough rectangle recognition algorithm, Contourlet-SIFT feature matching algorithm and HSV-Kmeans color segmentation algorithm.In addition, these algorithms have been experimentally verified and applied in industrial vision.For tile images with complex texture, the differences between foreground and background are small. As a result, conventional background difference algorithms fail to extract rectangular tiles correctly. The thesis uses the feature that Hough space is convenient for regular rectangle recognition, and then introduces Canny edge extraction and preprocessing of texture fill for optimization, so as to solve the problems in ROI extraction of tile images.For tiles whose texture is highly repetitive, the original SIFT algorithm suffers from serious mismatches and low efficiency. The thesis puts forward a Contourlet-SIFT invariant feature extraction algorithm, which combines local features with global features and combines space domain with time domain, to improve matching efficiency and reduce mismatches for regions that has similar texture. In that case, registration for tiles is implemented.For tiles that have multiple texture and color, it is difficult to judge the degree and region of color cast through a single threshold. The thesis extends Kmeans clustering algorithm to HSV color space and uses the results of color clustering to segment texture from base color. On that basis, unique threshold is set for each texture block to implement color cast detection.The innovations of this thesis are as follows:● A rectangle recognition algorithm in Hough space is optimized:edge detection and Hough transform are combined together with internal texture fill to extract tile ROI, which is easily interfered by irregular texture under complex backgrounds.● A Contourlet-SIFT invariant feature extraction algorithm is proposed:SIFT local invariant features and Contourlet global features are combined together to improve accuracy and efficiency of surface feature matching for tiles whose texture is highly repetitive and finally implement registration between tile images and templates.● Kmeans clustering algorithm is applied into image segmentation in HSV color space:HSV color space, Kmeans clustering and image segmentation technology are combined together to implement the segmentation of tile texture and background. It helps to set unique threshold and analyze the extent of color cast more accurately, qualitatively and quantitatively.