A Study of Image Analysis Techniques Based on Luminance/Color Contrast

• Main Authors: Hsin-Chia Chen, 陳信嘉
• Other Authors: Sheng-Jyh Wang
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/33055881755910767398

博士 === 國立交通大學 === 電子工程系所 === 95 === In this dissertation, a study of image analysis techniques by correlating subjective visual qualities with objective visual quantities based on luminance/color contrast is presented. To mimic the way humans perform image analysis, some subjective visual quantities are considered. To extract and verify the applicability of these visual quantities, subjective experiments are performed first. Then, to measure these subjective visual quantities, some objective quantitative measures based on luminance/color contrast are proposed. With these objective quantitative measures, contrast-based image analysis techniques can be developed for various image analysis applications. In the flow chart of a conventional image analysis system, four basic parts are included: 1) inputting of images to be analyzed, 2) image analysis with one or more techniques, 3) outputting of analyzed results, and 4) evaluation of the analyzed results. Specifically, given one or more images to be analyzed, different image analysis techniques are adopted for different applications. Then, the analyzed results are evaluated with some evaluation methods according to predefined visual perception requirements. In this dissertation, two more processes are added into an image analysis system. They are 1) subjective experiments and 2) measurement of luminance/color contrast and/or measurement of visual perception quantities. To mimic the way humans perform image analysis, we need some suitable subjective visual quantities. To extract appropriate visual quantities that may well correspond to humans’ perception, subjective experiments are needed. To estimate these subjective visual quantities for different applications, we need to propose effective and efficient objective quantitative measures. In this dissertation, we consider two different image analysis applications: 1) automatic inspection for visual defects on LCD panels, and 2) color segmentation. For different image analysis applications, the applicable visual quantities will be different. In the automatic defect inspection application, we discuss the suitable visual quantities for the extraction of visual defects with low luminance contrast. Here, we follow Mori’s proposal to quantify the degrees of image defects based on the luminance contrast and area size of visual defects. Based on Mori’s subjective experiments, which were performed to relate human visual perception with the luminance contrast and area size of visual defects, and the SEMU formula, which was proposed by Mori et al for a quantitative measurement of visual perception, we may effectively quantify the degrees of image defects based on luminance contrast and defect area. The LOG operator is then used to detect several types of visual defects. An optimal thresholding mechanism is also discussed. For the applications of color segmentation with little texture, we consider segmentation quality, degree of over-segmentation, and degree of under-segmentation as the visual quantities. To verify the correlation among these visual quantities, a few subjective experiments are performed. Here, we use color contrast to quantify these visual quantities. Usually, given a color image, adjacent pixels with low color-contrast are grouped into regions; while adjacent pixels with high color-contrast are regarded as edges. For color segmentation, we define color-contrast in terms of visible color difference and invisible color difference. Then, some objective quantitative measures based on visible/invisible color difference are proposed to measure these aforementioned subjective visual quantities. In this dissertation, the “intra-region visual error” is proposed to measure the degree of under-segmentation, while the “inter-region visual error” is proposed to measure the degree of over-segmentation. With these visual measures, some image analysis techniques are proposed to perform color segmentation and also the evaluation of color segmentation. With simulations for these two image analysis applications, some conclusions are drawn. First, the correlations between the luminance/color contrast-based quantitative measures and the visual quantities are really significant. Second, luminance/color contrast may play an important role in the development of image analysis techniques that mimic the way of human perception.