| 题名 | 内容相关的图像合成 |
| 作者 | 吴富章 |
| 学位类别 | 工学博士 |
| 答辩日期 | 2015-05-28 |
| 授予单位 | 中国科学院大学 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 张晓鹏 ; 董未名 |
| 关键词 | 图像合成 颜色转移 纹理合成 智能缩放 过程建模 形状语法 Image Synthesis Color Transfer Texture Synthesis Intelligent Retargeting Procedural Modeling Shape Grammar |
| 其他题名 | Content-based Image Synthesis |
| 学位专业 | 计算机应用技术 |
| 中文摘要 | 随着手机、数码相机等移动设备的迅猛发展，人们如今可以随意地获取大量的图像或视频数据。相比于传统的文字和音频信息，图像和视频能够将信息表达得更为丰富和直观，因此也开始逐渐取代文字成为互联网媒体的主角。图像数据的大规模增长增加了人们编辑和处理图像的需求，而传统的图像处理技术往往非常底层，用户需要掌握较多的专业知识和技巧。因此，研究直观、简单易用的图像视频处理技术，成为了目前相关研究领域的热点。本文主要研究自然场景图像与建筑表面图像的颜色风格与内容的合成方法。合成方法主要针对用户在图像内容、风格、尺寸等方面提出的技术、心理要求进行处理。在合成之前预先对合成对象进行语义分析，利用分析得到的结果指导合成过程。具体地，本文主要的研究工作和贡献如下： 1. 提出了一种内容相关的颜色转移方法。传统的颜色转移技术没有考虑图像的场景和内容，图像的颜色容易在不同语义的区域之间转移。为了避免这类问题，本方法预先对图像的语义区域进行分割，通过在语义区域之间建立映射关系，保证图像的颜色在同类语义区域之间进行转移。 2. 提出了一种基于特征感知的自然纹理图像合成方法。本方法在纹理合成之前预先对自然纹理图像的视觉特征进行分析，然后将得到的特征形式化为纹理合成的特定约束，最后可以明显地改善纹理合成的结果。 3. 提出了一种基于内容分析的自然图像缩放方法。该方法在缩放自然场景图像时能够自动地分析出图像内的纹理区域和非纹理区域，针对纹理区域，本文采用了一种新的纹理合成方法对其进行合成缩放。实验结果表明，该方法在缩放自然场景图像时能够得到更美观和合理的缩放结果。 4. 提出了一种基于语法的建筑表面图像分析和合成方法。本文利用分解语法规则来分析和表示建筑表面图像的内容结构，提出了一种近似的动态规划算法，用于自动地从被标记过的建筑图像中抽取出一套语法规则。提取得到的语法可以用于建筑图像合成以及大规模城市场景建模。 |
| 英文摘要 | With the rapid development of mobile device (such as mobile phones, digital cameras, etc.), people are freely able to get plenty of images or videos nowadays. Compared to traditional text and audio information, visual media like image and video can express information in a richer and more intuitive way. Because of this, image and video are gradually beginning to replace the text as the protagonist of the Internet. Image editing and processing techniques are urgently needed since the visual data is growing massively. However, most of the traditional image processing technologies are low-level operations. The users need to have lots of professional knowledge and skills if they want to edit image well. Therefore, studying a intuitive and easy-to-use framework for image and video processing has become the research focus in the related fields. In this thesis, we focus on the synthesis techniques for the natural scene images and facades. More specifically, the proposed synthesis algorithms mainly deal with the image’s content, appearance and size. We perform a pre-semantic analysis for the image and use the analysis result to guide our synthesis process. The main contributions of this thesis include as follows: 1. We propose a content-based color transfer approach. Since the traditional color transfer methods don’t analyse the image’s content, the color styles are easily transferred between the unrelated regions. To avoid these problems, the proposed method first performs an semantic image segmentation and builds a mapping between image semantic regions. Based on the mapping, the color styles will be transferred in the region pairs with similar content. 2. We propose a feature-aware natural texture image synthesis algorithm. Our synthesis method analyses the visual feature first and formulate the feature as a synthesis constraint. The experiments show that our algorithm generates a better result than the previous work. 3. We propose a new content-aware image retargeting method. The method first segments the image into texture and non-texture regions separately. For the texture region, we present a new texture synthesis method to resize it. Our experiments show that the proposed algorithm has distinct advantages to resize the natural scene images. 4. We propose a grammar-based structure analysis and synthesis algorithm for facade images. Analyzing the facade’s structure with split grammar is really helpful for facade synthesis and modeling. We p... |
| 语种 | 中文 |
| 其他标识符 | 201218014629099 |
| 内容类型 | 学位论文 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/6721]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 吴富章. 内容相关的图像合成[D]. 中国科学院自动化研究所. 中国科学院大学. 2015. |
| 个性服务 |
| 查看访问统计 |
| 相关权益政策 |
| 暂无数据 |
| 收藏/分享 |
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