单幅含噪模糊图像盲复原方法研究
文献类型：学位论文
| 作者 | 孙士洁 |
| 学位类别 | 博士 |
| 答辩日期 | 2017-05-25 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 沈阳 |
| 导师 | 赵怀慈 |
| 关键词 | 图像复原 盲去卷积 模糊核 图像显著结构 低秩先验 |
| 其他题名 | Single Noisy and Blurred Image Blind Deconvolution |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 图像复原作为图像处理和低层视觉中基础性的研究问题，一直广受关注。它旨在从观测的降质图像中重构或恢复出原始清晰图像，提高图像数据在后续图像分析、图像理解中的使用价值。该技术出现至今已成功应用于工业生产、遥感图像、智能监控、医疗影像、刑事取证和军事目标识别等众多领域。因此，开展图像复原研究具有重要的理论意义和实用价值。 图像复原本质上是一类不适定的反问题，考虑图像噪声影响会使该问题变得更为复杂。近来，不含噪声或包含十分少量噪声的单幅图像复原取得了显著的进步。然而，大多数单幅图像复原方法依然对图像噪声十分敏感，不能很好地处理含噪模糊图像复原问题。本文研究以单幅含噪模糊图像盲复原为主线，主要提出了基于图像显著结构和基于低秩先验的空间不变模糊图像盲复原方法以及基于投影运动路径模糊(Projective Motion Path Blur, PMPB)模型的空间变化模糊图像盲复原方法，并开发了对应问题的有效算法。本文所作的主要工作和创新点如下： (1)提出了一种利用图像显著结构的单幅空间不变模糊含噪图像盲复原方法。首先通过BM3D滤波对输入含噪模糊图像进行降噪预处理，进而利用基于全变分(Total Variation, TV)模型的方法，从降噪模糊图像中提取图像显著结构，并运用梯度选择处理移除不利于模糊核估计的显著边缘；然后借鉴两阶段模糊核估计策略，先利用基于图像显著结构的模糊核估计方法给出初始模糊核，再通过迭代支持域检测(Iterative Support Detection, ISD)技术纠正初始模糊核估计偏差；最后通过稀疏先验约束的非盲去卷积方法完成最终的图像恢复。实验结果表明，与已有代表性方法相比，提出方法在合成和真实图像上都能更准确地估计出含噪模糊图像的模糊核，获得高质量的复原图像，有效地处理图像盲复原对噪声敏感问题。 (2)提出了一种MAP框架下基于低秩先验的单幅空间不变模糊含噪图像盲复原方法。首先在估计中间过程复原图像过程中利用低秩先验约束对中间过程复原图像中的噪声进行抑制；然后采用降噪后的中间过程复原图像估计模糊核，以获得更准确的模糊核估计。迭代上述两个操作得到最终准确、可靠的模糊核估计；最后根据估计的模糊核，通过非盲去卷积方法得到最终的清晰图像。实验结果表明，提出方法在合成和真实图像上的定量和定性评价都优于已有代表性方法。 (3)提出了两种基于PMPB模型的单幅空间变化模糊含噪图像盲复原方法。以空间变化模糊的PMPB模型代替空间不变模糊模型，第一种方法引入基于图像显著结构的空间不变模糊含噪图像模糊核估计方法，利用图像显著结构信息估计空间变化模糊核；第二种方法是引入基于低秩先验的空间不变模糊含噪图像模糊核估计方法，利用MAP框架实现空间变化模糊核的估计。在此基础上，设计了基于PMPB模型的非盲去卷积方法。此外，为降低基于PMPB模型的空间变化模糊盲复原过程的计算量和存储量，还采取了基于高效滤波流(Efficient Filter Flow, EFF)的空间变化模糊进行近似技巧。实验结果表明，提出方法对去除单幅模糊含噪图像中的空间变化模糊是有效的，且恢复的图像具有更好的视觉效果。 |
| 英文摘要 | Image restoration, which is a fundamental research problem in image processing and low-level vision, has long attracted extensive attention. It is aimed at reconstructing or recovering its original sharp version from a given degraded image, helping to improve the application value of the given image in subsequent image analysis and understanding. Since it appeared, this technology has been applied successfully in various areas, such as industrial production, remote sensing image, intelligent surveillance, medical imaging, criminal forensics and military target recognition, etc. Therefore, the research of image restoration is of great theoretical significance and practical application value. Image restoration is an essentially ill-posed inverse problem, and additional image noise makes it even more complicated. Recently, significant progress has been made in single image restoration when the input image is free-noise or contains very little noise. However, most of state-of-the-art single image deconvolution methods are still sensitive to image noise, resulting in failing to solve well the noisy and blurry image restoration problem. In this thesis, we focus on single noisy and blurry image blind deconvolution, mainly proposing two different blind deconvolution methods to remove space-invariant blur: the first one uses salient image structure, while the second one uses low rank prior; and the method to remove space-variant blur based on projective motion path blur (PMPB) model. Meanwhile, the algorithms used to solve the corresponding problems also have been developed. The main work and innovation are as follows: (1) A single blurry and noisy image blind deconvolution method for space-invariant blur is proposed, using salient image structure. First, we use denoising as a preprocess to remove the input image noise, and then compute salient structure of the denoised result based on the total variation (TV) model. We also apply a gradient selection method to remove those salient edges that have a possible adverse effect on blur kernel estimation, thus improving the robustness of blur kernel estimation. Next, we adopt a two-phase blur kernel estimation strategy to achieve an accurate kernel estimation by taking advantage of the blur kernel estimation method from salient structure and iterative support detection (ISD) technique. Finally, we choose to use the non-blind deconvolution method with sparse prior knowledge to attain the final latent image restoration. Experiment results on synthetic and real world data show that our proposed method produces more accurate blur kernels and higher quality latent images than previous representative approaches on noisy and blurry images, handling effectively the problem that image deblurring techniques are very sensitive to noise. (2) A single noisy and blurry image blind deconvolution algorithm for space-invariant blur is proposed, using alternating maximum a posteriori (MAP) estimation combined with low rank prior. First, when estimating the intermediate latent image, low rank prior is used as the constraint that is used for noise suppression of the restored image. Then the denoised intermediate latent image in turn leads to higher quality blur kernel estimation. These two operations are iterated in this manner to arrive at reliable blur kernel estimation. Finally, we choose to use the non-blind deconvolution method with sparse prior knowledge to achieve the final latent image restoration. Extensive experiments on synthetic and real world data manifest the superiority of the proposed method over state-of-the-art representative techniques, both qualitatively and quantitatively. (3) Two single noisy and blurry image blind deconvolution approaches for space-variant blur are proposed, based on PBPM model. When PMPM model for space-variant blur is employed instead of space-invariant blur model, one approach is to estimate space-variant blur kernel via salient image structure, introducing the idea of space-invariant blur kernel estimation from salient image structure for single noisy and blurry image; another approach is to obtain space-variant blur kernel estimation via MAP frame, adopting the idea of space-invariant blur kernel estimation based on low rank prior. On the basis, a non-blind deconvolution method based on PMPB model is also developed. Moreover, to reduce the computation and storage costs of the whole blind deconvolution for space-variant blur based on PMPB model, an approximation technique for space-variant blur, a.k.a efficient filter flow, is also applied. Experiment results demonstrate that our proposed method is effective in removing space-variant blur of single noisy and blurry image while gaining higher visual quality in image restoration. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 源URL | [http://ir.sia.cn/handle/173321/20536]  |
| 专题 | 沈阳自动化研究所_光电信息技术研究室 |
| 推荐引用方式 GB/T 7714 | 孙士洁. 单幅含噪模糊图像盲复原方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所. 2017. |
入库方式： OAI收割
来源：沈阳自动化研究所
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。