Chinese Journal of Computers   Full Text
Title	Global Texture Optimization Incorporating with Image Detail
Authors	XIAO Chun-Xia1),2) HUANG Zhi-Yong2) NIE Yong-Wei2) LIU Meng2) HE Fa-Zhi2)
Address	1)(State Key Laboratory of Software Engineering, Wuhan University, Wuhan 430072) 2)(School of Computer, Wuhan University, Wuhan 430072)
Year	2009
Issue	No.6(1196—1205)
Abstract &   Background	Abstract This paper presents a new global texture synthesis method incorporating with image detail. The image detail extracted using the no-linear method is used as a new channel for guiding the texture synthesis. Integrating the image detail with the color channels, the authors build a global energy function on the image, and the function is optimized using Expectation Maximization. They further improve the texture quality by applying image detail and color histogram matching techniques. Both image and video synthesis methods are presented in this paper. Experiments on a wide variety of image and video synthesis examples demonstrate the advantages of the proposed method over existing techniques: keep the synthesized texture continuously, and the broken structure of the texture are avoided. Keywords texture synthesis; global optimization; expectation maximization; bilateral filtering; image completion Background This work is partly supported by the National Natural Science Foundation of China (No.60803081), State Key Lab of CAD&CG (No.A0808), State Key Laboratory of Software Engineering(SKLSE) (No.SKLSE2008-07-08), Ph.D. Programs Foundation of Ministry of Education of China (No.200804861038), Natural Science Foundation of Hubei Province (No.2008CDB350). Texture synthesis is a useful technique in computer graphics, it synthesize a large scale texture based on a small sample while exhibiting the same stochastic features of the exemplar texture. A variety of methods have been developed to achieve this goal. Synthesis of novel photo-realistic imagery from limited example input is of wide importance in computer graphics. Many example based synthesis approaches rely on the presence of texture. Texture refers to the class of imagery that can be categorized as a portion of an infinite pattern consisting of stochastically repeating elements. This inherent repeatability present in textures is the key behind texture synthesis techniques. These techniques generate output textures that are larger in size than the input sample but perceptually similar to it. The traditional approach in texture synthesis is to compare an image patch with that of an exemplar. For textures with strong large structures, feature maps can be particularly helpful. It has been shown that some textures can be synthesized better with the aid of feature maps, which provide non-local feature information. In the authors’ method, different from the binary feature maps used in existing methods, the high frequency detail of the texture and its spatial variation are used to improve the synthesis results. By extracting the image detail using the no-linear method, the image detail is used as a new channel for guiding the texture synthesis. Integrating the image detail with the color channels, the authors build a global energy function on the image, and the function is optimized using Expectation Maximization. They further improve the texture quality by applying image detail and color histogram matching techniques. Both image and video synthesis methods are presented in this paper. They have also extended these techniques to allow for image and video completion based on the constrained texture synthesis. Compared with the latest texture synthesis method using optimization, the authors’ method demonstrates the following advantages over existing techniques: keep the synthesized texture continuously, and the broken structure of the texture are avoided. The result of existing methods employ color histogram matching looks more quite similar to the exemplar, but these methods are not good at maintaining the continuity of structural features as well as the shapes of individual objects in the textures. It shows that if the color of the exemplar is similar, the histograms of the existing methods also do not work well, while the detail-aware patch similarity matching makes better results.