Date of Award
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Frank Y. Shih
Yun Q. Shi
Textures have been used effectively to create realistic environments for virtual worlds by reproducing the surface appearances. One of the widely-used methods for creating textures is the example based texture synthesis method. In this method of generating a texture of arbitrary size, an input image from the real world is provided. This input image is used for the basis of generating large textures. Various methods based on the underlying pattern of the image have been used to create these textures; however, the problem of finding an algorithm which provides a good output is still an open research issue. Moreover, the process of determining the best of the outputs produced by the existing methods is a subjective one and requires human intervention. No quantification measure exists to do a relative comparison between the outputs. This dissertation addresses both problems using a novel approach. The dissertation also proposes an improved algorithm for image inpainting which yields better results than existing methods.
Firstly, this dissertation presents a methodology which uses a HSI (hue, saturation, intensity) color model in conjunction with the hybrid approach to improve the quality of the synthesized texture. Unlike the RGB (red, green, blue) color model, the HSI color model is more intuitive and closer to human perception. The hue, saturation and intensity are better indicators than the three color channels used in the RGB model. They represent the exact way, in which the eye sees color in the real world.
Secondly, this dissertation addresses the issue of quantifying the quality of the output textures generated using the various texture synthesis methods. Quantifying the quality of the output generated is an important issue and a novel method using statistical measures and a color autocorrelogram has been proposed. It is a two step method; in the first step a measure of the energy, entropy and similar statistical measures helps determine the consistency of the output texture. In the second step an autocorelogram is used to analyze color images as well and quantify them effectively.
Finally, this disseratation prsesents a method for improving image inpainting. In the case of inpainting, small sections of the image missing due to noise or other similar reasons can be reproduced using example based texture synthesis. The region of the image immediately surrounding the missing section is treated as sample input. Inpainting can also be used to alter images by removing large sections of the image and filling the removed section with the image data from the rest of the image. For this, a maximum edge detector method is proposed to determine the correct order of section filling and produces significantly better results.
De, Chandralekha, "Example based texture synthesis and quantification of texture quality" (2012). Dissertations. 324.