模糊图像复原若干技术研究
文献类型：学位论文
| 作者 | 段江永 |
| 学位类别 | 工学博士 |
| 答辩日期 | 2014-04-30 |
| 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 | 潘春洪 |
| 关键词 | 图像去模糊 图像反卷积 区域约束 多焦融合 振铃去除 image deblurring image deconvolution region constraint multifocus image fusion ringing removing |
| 其他题名 | Research on restoration of blurred images |
| 学位专业 | 计算机应用技术 |
| 中文摘要 | 在数字图像处理中，模糊图像复原具有重要的研究意义和应用价值。由于模糊图像复原是一个不适定问题，使得从模糊图像中恢复清晰图像非常困难。虽然目前大量算法已被提出，但仍然存在一些不足，比如恢复的图像中仍存在一些模糊及振铃。在多焦图像融合复原中，已有算法也存在一些振铃、区域重叠以及虚假边缘等现象。本文提出了几种新模型，以解决目前方法中的这些问题。本文的主要贡献点如下： 1. 提出了一种基于图像相似性去除失焦模糊的方法。基于模糊图像的某些区域通常在相似清晰图像中存在对应的清晰区域的观测，算法分别从模糊图像和相似清晰图像中通过SIFT 配准提取区域块对，然后根据区域块对之间的约束关系估计出模糊核，最后通过反卷积算法恢复出清晰图像。该算法证明了相似清晰图像可以提供有用信息用于去模糊，从而提供了一种有效利用其他图像辅助去模糊的新途径。 2. 提出了一种基于区域融合的多焦图像复原方法。论文把多焦图像复原问题看作区域分割融合的问题，并对区域分割提出了三个直观假设，然后在这些假设的基础上建立了关于区域分割的目标函数。目标函数包含了实现无缝重构的重构误差项、去除振铃的模糊度量项以及使图像中物体分割一致的平滑先验项。这三项共同显式地对多焦融合中存在的人工痕迹进行建模。此外，论文还提出了一种贪婪算法优化目标函数。算法可以得到没有明显虚假边缘与振铃等人工痕迹的清晰图像。 3. 提出了一种基于边缘区域约束下的直线运动去模糊复原方法。论文推导出了运动模糊核和模糊边缘区域之间的约束关系，并利用该约束估计运动模糊核。同时，提出了一种模糊图像边缘区域自动获取算法，并且采用RANSAC算法对获取的边缘区域进行有效筛选，去除了外点边缘区域的影响，使模糊核估计算法更加鲁棒准确。 4. 提出了一种多尺度运动模糊图像复原算法。算法在图像的多个尺度上依次迭代求解清晰图像与模糊核，并在得到最小尺度上的模糊核后，通过反卷积算法恢复出清晰图像。论文估计了图像中的平滑区域，建立了平滑性区域约束，可有效去除这些区域中的振铃，同时使得模糊核的估计更加准确；此外，在反卷积算法中建立了融合图像噪声先验、 图像统计先验、图像稀疏表示先验以及平滑区域约束的模型，进一步抑制了图像中振铃的产生。 |
| 英文摘要 | In the field of image processing, image deblurring is one of the most important technologies. It is of great important value in both application and research areas. As a challenging ill-posed problem, many image deblurring methods have been proposed to restore the clear images. However, the blur and ringings still exist in these methods. For multifocus image fusion, the existing methods suffer from ringing, overlapping and artificial edges. In this thesis, we develop new models to solve these problems. The main contributions of this thesis are as follows: 1. A method is developed to remove the out-of-focus blur based on the similarity between images. It arises from the observation that a patch in a blurred image can often have a corresponding clear one in their similar images. Our method first extracts the patch pairs by using SIFT matching. Then, we estimate the blur kernel by the constraint between the patch pairs via the RANSAC algorithm. Finally, the clear image is restored by image deconvolution. We demonstrate that similar clear images can provide useful information for image deblurring, which provide a new way to efficiently restore the blurred image with the use of additional images. 2. A new framework is developed to restore multifocus images based on region fusion. Considering the multifocus fusion as the region segmentation, some hypothesises concerning with the segmentation are first proposed. Then we enforce these hypothesises by developing a unified optimization framework. The objective function of the framework contains three terms, i.e., the reconstruction error to realize seamless region reconstruction, the out-of-focus energy to remove the ringings and the smoothness prior to enforce the consistent segmentation for the objects in the image. These three terms together explicitly model the common visual artifacts in the fusion process. We then propose a greedy algorithm to minimize the objective function. Our approach can produce a clear image with few artificial edges and ringing artifacts. 3. An algorithm is proposed to remove uniform linear motion blur. Based on the observation that clear images often have step edges, we develop a new edge region constraint. With this type of constraints, the motion kernel is efficiently estimated. Besides, we propose a method to automatically extract the edge regions. Some outlier regions are then rejected by using the RANdom SAmpling Consensus (RANSAC) algorithm, which... |
| 语种 | 中文 |
| 其他标识符 | 200818014629087 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/6577]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 段江永. 模糊图像复原若干技术研究[D]. 中国科学院自动化研究所. 中国科学院大学. 2014. |
入库方式： OAI收割
来源：自动化研究所
浏览0
下载0
收藏0
其他版本
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。