Image Analogy

Abstract

Texture synthesis is a popular topic in image processing in recent years. Its applications are commonly seen in various fields, such as computer graphics and computer vision. Image analogy is one of the special applications of texture synthesis, which can transform textures by examples. However, generating high quality images with large size by texture synthesis usually takes long running time. It is crucial to have a balance between efficient and high quality work. In this project, the algorithm of texture synthesis and its applications were investigated. Among those different synthesis approaches, the algorithm derived from Markov Random Field was selected for investigation. This algorithm allows the output images generated through a searching processing. It can preserve the original structure of a broad range of texture, inclusive of realistic textures. However, the traditional method for texture synthesis requires full search for each pixel, which is extremely time consuming. To reduce the high computational cost of texture synthesis, three types of partial search algorithms were proposed to speed up the process of texture synthesis. They are Diamond Search, Cross Search and Hexagon Search. The results generated from the three partial search algorithms were compared to examine how accelerators affected the results in terms of quality and time cost.