A Study of Efficient Techniques for Texture Synthesis and Image Transfer

• Main Author: 張任烜
• Other Authors: 王宗銘
• Format: Others
• Language: en_US
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/37662405739631567088

碩士 === 國立中興大學 === 資訊科學研究所 === 90 === Texture synthesis has been the subject of intensive research for many years in computer graphics community. Recently, Efros addressed this topic by presenting an algorithm referred to as image quilting, where a new image is synthesized by stitching together small patches of existing images. Efros also extended the algorithm to perform texture transfer, rendering an object with a texture taken from a different object. Efros’s approach produces good results for a wide range of textures. However, it demonstrates a severe drawback in the execution time, requiring several minutes for quilting and many iteration processes for image transfer. Another disadvantage of his method is that the synthesized results are normally with sharp edge in overlapped patch boundaries. In this thesis, we present a number of techniques to improve such drawbacks. We also provide several applications of texture synthesis and image transfer that have not been addressed by Efros. In particular, we utilize a new search algorithm to determine the best match texture patch to improve the computing performance. We present a technique, the minimum error path, to calculate the path position with the minimum but constraint variation in the overlapped patch boundaries. Apart from applying this technique, we blend the overlapped region with the weighted linear feathering method we proposed to reduce the visual perception of sharpness. For the image transfer, we present a novel algorithm, non-iterative image transfer algorithm. This algorithm utilizes the feature vectors constructed in the overlapped patch according to the order from the left and bottom region as well as from the right and top one. Given the feature vectors, our algorithm then efficiently determines a patch that is best match using the K-d tree search algorithm. We implemented our method as well as Efros’ approach, and collect some experimental results for comparison and analysis. We discover that our method works remarkably well, producing results that are better than Efros’ approach. In addition, our method significantly reduces a fraction of the computational cost, which requires not many but only a single iteration process for image transfer. Furthermore, our method is nearly 42.91% faster than our counterpart under the comparison on an iteration process. Finally, we present several applications of texture synthesis and image transfer, including hole filling, texture extrapolation, and color transfer. We conclude that this research has proposed efficient approach for texture synthesis and image transfer, and it has extended their feasibility by presenting several applications with very promising results.