Color Image Segmentation Approaches For Printing Quality Inspection

Inspection of printing quality is an important task in modern printing industry, and directly influences the quality and productivity of the printing process. With the printing process being highly nonlinear, uncertain and time-varying, traditional inspection theory and simple image processing techniques have limited capability in online inspection of multi-objects across the whole printed picture. This paper presents a new printing quality inspection approach based on color image segmentation techniques and an evaluation system consistent with human visual perception that bridges the gap between the subjective and object evaluations on printing quality. Major contributions of this paper include:(1) An image quality evaluation model based on human visual perceptionConsidering the characteristics of printing technology and human visual perception, printing quality can be divided into image quality and text quality. A basic printing image quality evaluation framework based on human visual perception is therefore proposed that includes color measurement, image inspection, and automatic control. Inside the image inspection, text localization, image ROI (region of interest) segmentation, and half-tone inverse transform are the three key techniques that help to establish an objective printing quality evaluation and control system that is multi-object, multi-mode and consistent with subjective evaluation.(2) An on-line color measurement method based on illuminating color compensationSince the lighting condition for printing quality inspection may not meet the standards, an illuminating color compensation method is designed to deal with this variability. Dichromatic reflection model is used to estimate the spectral power distribution of the illumination, and surface reflectance functions is reconstructed from received color signals, based on which colors in the image are able to be recovered using the finite-dimensional model. An improved color difference calculation method on RGB color space is also proposed as well, to increase the computation speed without losing accuracy. The results show that this on-line color measurement system can meet the requirement of industry standards with high accuracy and speed. (3) The intelligent measurement model for the percentage of the area covered by cyan, magenta, yellow and black inks of printed half-tone pictureCalculating the percentage of the printed CMYK ink dot area using a simple model is hard due to the complexity of the printing process and the influence of spot-color and black ink technique. A complex RGB-CMYK transform strategy is therefore proposed based on ensemble neural network on primary color regions. Since similar colors have similar percentages of printed dot ink area, an adaptive color image segmentation algorithm is designed to reduce the number of colors using fuzzy sets 2 and fuzzy homogeneity techniques. Experimental results show that the method can achieve higher accuracy of the percentage of the printed CMYK ink dot area, and hence help considerably in evaluating the color ink printing quality and realizing automatic control for printing production.(4) A text region localization algorithm based on the primary color plane and connected component featuresFor the difficult problem of text region localization in complex color images with different languages, fonts and colors, an algorithm based on primary color plane and connected component features is proposed. The local smoothed colors at the interior of strong edges are selected to be seeds of clustering algorithm for coarse segmentation, which help to retain the color of text in the picture. The mean shift algorithm is used to further reduce the number of primary colors as a refined segmentation to enhance the speed of calculation. And a maximum likelihood filtering technique is used as a post-processing to increase the recognition accuracy. This filtering, combined with heuristics, region-based geometric location information and a reconstruction rule system that uses expert knowledge based on the structural features of Chinese characters, can successfully segment text regions from the printed color picture. Experiments show that the method can effectively remove the interference from complex backgrounds and adapt to different fonts, scales and other factors in various types of images with precise text locations and successful text/image segmentations obtained.(5) An image ROI (regions of interest) segmentation algorithm based on human visual perceptionAiming at achieving the consistency of printed image quality inspection with human visual perception, an improved color and texture-based segmentation method is designed. Since human visual cortex can be decomposed into a series of independent channels of orientation and spatial frequency, pyramid-based texture segmentation method is used to divide the image into six types of texture regions. The optimal compositional color distance algorithm is applied to each region to extract the color features that embodies not only the spatial context but also the local and global information of image consistent with the human visual criteria. In each texture region, an adaptive mean shift algorithm is then used, combining color and texture features, to segment the picture into sub-regions with similar texture and color. Those sub-regions, when further merged, correspond to different objects and backgrounds in picture. Experimental results show that the algorithm is robust against illumination variations, and the obtained ROIs have similar color, texture and semantic features, laying a solid foundation for setting up a human visual perception based image quality evaluation system.