Acceleration Strategy Of Regular Patch-based Texture Synthesis

Texture synthesis is a way to simulate and generate the pattern of texture’s appearance. It is one of the hot issues in the research field of computer graphics, image processing and computer vision in recent years. Within the methodology of example-based texture synthesis, patch-based algorithm has a very wide range of applications, while the research of regular patch-based texture synthesis occupies a large proportion in it.This article is mainly focus on the acceleration strategy of regular patch-based texture synthesis. After the introduction of related concepts and the summary and analysis of the previous works, two main research aspects of this paper are addressed as follows:1. Task scheduling for regular patch-based texture synthesis in the parallel wavefront pattern. The algorithms based on patch are hard to be parallelized due to their relatively high constraint between adjacent blocks. This article follows one of the classic algorithms, transforming the problem of finding optimal synthesis path into a task scheduling problem, describing and extracting constraints from the problem, and then a spatial balanced schedule based on wavefront pattern is proposed in order to deal with the general2d grid data. After that we put forward a scheduling algorithm for regular patch based texture synthesis which keeps the original constraints well, and perform our implement on a multi-core CPU. Analysis and experiment results are then given to prove the effectiveness of the solutions.2. Fast near-regular texture synthesis based on the concept of co-evolution. Texture evolutionary system is a near-regular texture oriented synthesis algorithm. The main feature of the system is the definition of behaviors which on the basis of the evolution theory to minimize the error accumulate problem caused by the inflexibility of synthesized blocks. Inspired by previous work, this paper presents a new texture synthesis algorithm on the concept of co-evolution. Behaviors are modified deeply for the main purpose of obtaining better efficiency. The approach of pre-computation reduces the compute redundancy and the new process of evolution can be better parallelized. Results show that the proposed algorithm does not only reinforce the versatility but also improve the efficiency greatly.