紋理合成技術在透視失真紋理合成及視覺風格轉換的研究

The Research of Texture Synthesis for Perspective Distorted Texture and Visual Style Transfer

鄭明政(Ming-Cheng Cheng)

指導教授 ： 郭忠民

義守大學/電機資訊學院/資訊工程學系/博士(2016年)

摘要

透視失真的紋理普遍存在真實世界中，這一類型的紋理大部份是人為因素所導致，例如攝影時沒有垂直面對景物拍攝就會產生有透視感的影像，因此產生了透視失真的紋理，傳統以區塊為基礎的紋理合成法並不適用於這類型的紋理, 在這份研究中我們將建構一個自動修正模組，使用這一修正模組自動修正透視失真的紋理，然後再進行紋理合成，同時也會改進某些以區塊為基礎的紋理合成法的缺點，使用混合合成法來改進區塊合成時邊界的不連續性，並使用修改過的粒子群聚法(PSO)來增進計算效率。
以往的影像視覺風格轉換法並非完全自動，太依賴個人主觀想法，因此發展一個可以自動完成影像視覺風格轉換並高度符合使用者需求的影像視覺風格轉換演算法是必要的，我們將發展一有效的方法，對區塊大小做activity guided分析，找出最適合做視覺風格轉換的區塊大小，使用混合合成法來改進區塊合成時邊界的不連續性，並使用修改過的粒子群聚法(PSO)及區塊配對法來增進合成效率。

關鍵字

紋理合成 ； 視覺風格轉移

並列摘要

Perspective distorted textures are pervasive in man-made and real world, which are usually captured by camera lens with non-vertical camera axis. However the conventional texture synthesis approaches cannot be applied to non-parallel textures. In this research, we construct a control structure correction model based on an automatic rectification approach to synthesize the appearance of the perspective distorted textures faithfully. Furthermore, some common drawbacks of patch-based sampling algorithm such as blending procedure and high computational cost will be improved. In this research, we develop a hybrid blending scheme to improve visual transition artifacts along patch boundaries, and also apply Particle Swarm Optimization (PSO) to improve the computational efficiency.
The research of image visual characteristics is motivated from texture synthesis techniques. Generally, the success of these visual style transfer applications is inherently user oriented and highly depends on the subjective personal thoughts also. Therefore, a method with the reasonable estimation values to synthesize a styled texture which would meet the user preferences is highly desirable. In this research, the algorithm starts with an activity guided analysis for adaptive patch-based visual characteristics transfer, which is not only more efficient than conventional methods but also with pleasing visual quality. Then, the Particle Swarm Optimization (PSO) accelerated scheme with a modified match criterion which can effectively search and the approximate best location that matches the synthesized target patch according to user-specified features. In addition, a hybrid blending approach ensures a Coherence Match to improve the transition effect between the overlapping boundaries of adjacent patches.

並列關鍵字

Texture Synthesis ； visual style transfer

參考文獻

[1] K. Sangeetha1, Dr. P. Sengottuvelan and E. Balamurugan, “Improved Exemplar Based texture Synthesis Method for Natural Scene Image Completion, “ International Journal of Computer Science Issues, Vol. 8, Issue 5, No 2, September 2011.

[2] A. Shah, P. Bhatt and K. J. Modi, “Image Enhancement Techniques by Texture Synthesis, “ International Journal of Emerging Technology and Advanced Engineering, Volume 2, Issue 1, January 2012.

[4] Ankitha A, “Exemplar Based Image Inpainting,“ Journal of VLSI and Signal Processing, Vol 4, Issue 2, pp. 13-16, Mar-Apr. 2014.

[5] H. S. Lee, “A New Image Super Resolution by Texture Transfer,” IEEE International Conference on Image Processing, pp. 3915-3918, 2014.

[6] U. S. Thakur and O. Chubach, “Texture Analysis and Synthesis using Steerable Pyramid Decomposition for Video Coding, ” IEEE International Conference on Image Processing, pp. 204-207, 2015.