Acceleration and Applications of Texture Synthesis

Abstract

Texture synthesis is vital for variety of applications in image processing and computer vision. However, in most texture synthesis algorithms, the synthesis speed is very slow. In this project, the Wei and Levoy's algorithm, and the Efros and Leung's algorithm are studied. Both of them are derived from the Markov Random Field texture model. I implement these two algorithms and compare their performance. Both of them are able to produce good visual quality in texture synthesis. However, in the Wei and Levoy's algorithm, the window mask is restricted to L-shape. Hence, the Wei and Levoy's algorithm is not appropriate for handling some applications, such as image inpainting and image extrapolation. Also, the computational complexity of these two algorithms is very high. The Efros and Leung's algorithm is then selected in this project to propose an acceleration version based on the k-means clustering. Besides, several applications, such as, image inpainting, image extrapolation, foreground removal and image analogy, are implemented.