Date of Award
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Frank Y. Shih
James A. McHugh
Yun Q. Shi
The focus of this dissertation is mainly on color image, especially on the images with lossy compression. Issues related to color quantization, color correction, color image retrieval and color image quality evaluation are addressed. A no-reference color image quality index is proposed. A novel color correction method applied to low bit-rate JPEG image is developed. A novel method for content-based image retrieval based upon combined feature vectors of shape, texture, and color similarities has been suggested. In addition, an image specific color reduction method has been introduced, which allows a 24-bit JPEG image to be shown in the 8-bit color monitor with 256-color display. The reduction in download and decode time mainly comes from the smart encoder incorporating with the proposed color reduction method after color space conversion stage. To summarize, the methods that have been developed can be divided into two categories: one is visual representation, and the other is image quality measure.
Three algorithms are designed for visual representation:
(1) An image-based visual representation for color correction on low bit-rate JPEG images. Previous studies on color correction are mainly on color image calibration among devices. Little attention was paid to the compressed image whose color distortion is evident in low bit-rate JPEG images. In this dissertation, a lookup table algorithm is designed based on the loss of PSNR in different compression ratio.
(2) A feature-based representation for content-based image retrieval. It is a concatenated vector of color, shape, and texture features from region of interest (ROI).
(3) An image-specific 256 colors (8 bits) reproduction for color reduction from 16 millions colors (24 bits). By inserting the proposed color reduction method into a JPEG encoder, the image size could be further reduced and the transmission time is also reduced. This smart encoder enables its decoder using less time in decoding.
Three algorithms are designed for image quality measure (IQM):
(1) A referenced IQM based upon image representation in very low-dimension. Previous studies on IQMs are based on high-dimensional domain including spatial and frequency domains. In this dissertation, a low-dimensional domain IQM based on random projection is designed, with preservation of the IQM accuracy in high-dimensional domain.
(2) A no-reference image blurring metric. Based on the edge gradient, the degree of image blur can be measured.
(3) A no-reference color IQM based upon colorfulness, contrast and sharpness.
Fu, Yan-Yu, "Color image quality measures and retrieval" (2006). Dissertations. 745.