A Texture Transfer Algorithm with Characteristics Preservation

• Main Authors: Kuo-Chen Wu, 吳國禎
• Other Authors: 王宗銘
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/64111250120428907568

碩士 === 國立中興大學 === 資訊科學系所 === 94 === Texture transfer is an important research subject in computer graphics community. Given a source texture and a target image, the texture transfer algorithm produces a result image containing the features of both the source texture and the target image simultaneously. Current texture transfer algorithms have two drawbacks. First, it is incapable to retain the structural patterns of the source texture as well preserving the detailed characteristics of the target image. Second, the algorithm is inefficient. In this thesis, we present an efficient and detail-preservation texture transfer algorithm to solve the above drawbacks. Our algorithm is a patch-based approach, operating the texture transfer patch by patch in the current processing block. Every texture-transferred patch is later pasted up on the target image with an overlapped region to maintain the texture continuity. Our algorithm consists of three steps including the variance detection, the adaptive patch sampling, and the patch replacement. In the first step, we calculate the variance of pixels in the processing block, and then compare it with a user defined threshold to determine whether the processing block contains the subtle characteristics of the target image. In case of containing the subtle characteristics, in the second step, we subdivide the block into smaller sub-blocks and repeat the variance detection step, until the subdivision has reached the termination criteria. In the third step, if no further subdivision is required, we search the kd-tree nodes constructed from a number of patches in the source texture, finding the best-matched patch with respect to the overlapped region in the processing block. Once found, we temporarily subdivide this patch into smaller sub-patches, on which we operate the texture transfer. Finally, the patch being texture transferred is pasted up on the processing block in the target image. We repeat the above three steps until all blocks have been operated the texture transfer completely. We implemented our algorithm and evaluated the performance using textures in the public texture database. We also compared our results with those presented by Efros’，Hertzmann’s, and Ashikhmin’s. The experimental results show that our algorithm is superior to Hertzmann’s and Ashikhmin''s algorithms in preserving the structural patterns of the source texture. Also, our algorithm can retain characteristics of the target image that has visual appearance better than our counterparts. In addition, the efficiency of our algorithm is higher than that of Hertzmann’s and Ashikhmin''s methods, which are the pixel-based approach. Finally, our algorithm can operate the texture transfer within several minutes using a single execution process. In contrast，Efros’ method requires several iteration processes, which takes longer time to produce the final result. In conclusion, our algorithm effectively solves the drawbacks of the current texture transfer algorithm. Our algorithm can preserve the structural patterns of the source texture as well as retaining the characteristics of the target image. The algorithm performs fast, requiring no iteration, and produces synthesized results in a short time. We consider our algorithm makes a concrete contribution to the texture transfer studies.