Image quality evaluation in print and compressed images

The digital image quality degrades during the process of capture, storage, transmission, display, or printing. Image quality evaluation plays an important role in digital imaging products design and development, e.g. printers, displays, scanners, and cameras. In this thesis, we contribute two tools to help understanding and evaluating how customers perceive image quality. In the first part, we study the print artifact banding from the aspect of artifacts characterization and simulation. Banding remains to be a significant print artifact that will dominantly influence the perceptual print quality for high-end digital presses. It is perceived as a singular or periodic fluctuation in the paper process direction. As digital presses place one color separation at a time, the contrast and spatial pattern of the artifacts depend on the local point within the color space about which the defect variation occurs. An innovative framework of banding characterization in full color space is developed to capture these color-dependent banding features. The simulation framework of banding artifacts on natural document images is proposed, in which a single banding prototype is used to ensure the spatial structure on the whole page and the simulated artifacts variations features are modulated according to the background colors. The proposed software simulator is a necessary tool for the development of automated artifacts detection and identification tools, and print quality assessment tools which could predict artifacts visibility for a given print content. In the second part, we investigate the visibility of ringing artifacts due to JPEG compression, from the aspect of developing an artifacts visibility predictor based on psychophysics experiments. JPEG ringing artifacts refer to the noisy surroundings of major edges that are supposed to have a smooth background. A novel psychophysics experiment is designed to collect human perception responses in local regions of an image containing compression artifacts, in which observers used a graphics tablet and pen to directly mark the displayed image, containing compression artifacts, by indicating both the location and the strength of those artifacts. A set of image features extracted from each unit segment of the potential ringing regions and the corresponding subjective ringing visibility ranking are used to train and test a non-reference ringing visibility predictor based on machine learning frameworks. The experiment and predictor development framework proposed herein will be a valuable resource for compressed image quality evaluation and ringing artifacts reduction preserving high-frequency texture