| 纹理优化合成的加速算法 |
| 黄浩达
|
| 2007-06-05
|
| 学位授予单位 | 中国科学院软件研究所
|
| 学位 | 博士
|
| 学位授予地点 | 软件研究所
|
| 关键词 | 纹理合成
能量最小化
并行化
图形处理芯片
流场可视化
|
| 摘要 | 纹理合成技术能基于小纹理生成视觉上相似的大纹理，它在计算机图形学及计算机视觉中有很广泛的应用。例如，大规模场景的绘制通常需要大量的纹理来丰富其细节表达，如果直接使用大纹理会给外存和内存带来很大的负担，而应用纹理合成技术可以改善绘制时的存储问题；在科学计算可视化方面，纹理合成技术也能用来提高相关内容的表现力，如流场的可视化。
由于纹理合成技术的重要性，近年来有大量的工作致力于改进纹理合成技术的速度、质量及合成能力。纹理优化是其中一种新颖的合成方法，它将纹理合成表示成能量最小化问题，并通过类期望最大化算法来逐步优化目标纹理。它不但能合成高质量的纹理，而且可以应用于带约束的纹理合成，如流场的纹理合成。但是，相比于其它合成方法，纹理优化方法的合成速度较慢，限制了它的实践应用。
基于纹理优化的思想，本文提出一种可运行于图形处理单元（GPU）的并行纹理合成算法，能很好地提高合成速度，且不降低合成质量。在新算法中，我们首先引进k-coherence search 和主元分析(Principle Component Analysis)技术使得纹理优化可在GPU上并行计算，然后我们利用局部像素的连续性和局部邻域的合成稳定性设计了两个加速技术，进一步提高了纹理合成速度。虽然新算法需要一定的预计算，但相关的开销不大。经对比实验，新算法的在线合成速度能比2005年ACM SIGGRAPH大会上提出的初始的纹理优化算法快4000倍以上，能满足交互式应用的要求。文章最后，我们用新算法实现了纹理流场的交互式编辑。 |
| 其他摘要 | Texture synthesis aims at generating visually similar large textures from a small exemplar texture，and has many applications in computer graphics and computer vision. For example, in using textures to enrich details of large scale scenes, texture synthesis can greatly reduce storage requirements in comparison with using large textures directly. In scientific visualization, texture synthesis is also always used to improve the visualization efficiency, such as for visualization of vector fields.
Due to the importance of texture synthesis, there have been many methods proposed to increase the synthesis speed, quality and ability of texture synthesis in recent years. Among these methods, texture optimization is very interesting as it formulates texture synthesis as an energy minimization problem, and so produces target textures by iterative refinement through a EM-like algorithm. It’s able to synthesize high quality textures, and more over, it could be applied to constrained texture synthesis such as flow-guided synthesis. However, its synthesis speed is slower than other methods. This limits its use in many interactive or real time applications.
In this paper, a novel method is proposed to speed up texture optimization without lowering the synthesis quality. At first, k-coherence search and principle component analysis (PCA) are used to make texture optimization be able to implement in parallel on graphics processing units (GPUs), then two acceleration techniques are developed by two features of texture optimization, local pixel coherence between neighbour pixels and local neighbourhood stability in the iteration process. With a reasonable cost on precomputation, the new method can synthesize textures 4000+ times faster than the original texture optimization method proposed in SIGGRAPH 2005, using about 77~87ms to produce a 256*256 textures from a 64*64 examplar texture. Thus, our new method is capable of interactive applications. Benefited from the high synthesis speed, we apply the new method to interactive editing of flow-guided synthesis. |
| 页数 | 45
|
| 语种 | 中文
|
| 内容类型 | 学位论文
|
| URI标识 | http://ir.iscas.ac.cn/handle/311060/5580
|
| 专题 | 中科院软件所_中科院软件所
|
推荐引用方式
GB/T 7714 |
黄浩达. 纹理优化合成的加速算法[D]. 软件研究所. 中国科学院软件研究所,2007.
|
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。
修改评论