【作者】 丁静；

【导师】 王培康；

【作者基本信息】 中国科学技术大学 ， 通信与信息系统， 2011， 硕士

【摘要】 超分辨率重建(Super Resolution Reconstruction, SRR)是图像处理领域的研究热点之一,其核心思想是利用一系列非冗余的低分辨率图像重建一幅高分辨率的清晰图像,本质上是一种用时间带宽换取空间分辨率的信息融合技术。本文首先介绍了超分辨率重建的基本原理、退化模型、正则化SR重建算法以及重建质量评价。通过比较分析现有的一些重建算法,发现算法的鲁棒性是衡量其能否进入实用化阶段的一项关键要素。对于现有常用的基于最小二乘估计的正则化SR重建算法而言,由于重建过程中可能会存在一些异常数据,导致估计结果偏差较大、算法鲁棒性差。本文深入地研究了采用M-估计的鲁棒正则化SR重建算法,并从两方面讨论了M-估计的选取,指出在正则化SR重建算法中采用Huber估计不仅可以保证算法的鲁棒性、解的唯一性,且获得的估计结果偏差较小。另外,引入稳健估计理论中的绝对中位偏差(Median Absolute Deviation, MAD)方法来计算M-估计所需的尺度参数,避免了人工选取带来的偏差和鲁棒性损失的影响。通过对比分析,实验结果验证了这种基于Huber估计的正则化SR重建算法的鲁棒性和有效性。对于SR重建,正则项不仅可以加速算法的收敛和提高解的稳定性,且直接影响图像处理的视觉效果,如对图像的边缘、纹理等几何结构的保持。如何设计合理的正则项来获得较好的边缘保持特性和视觉效果,一直是SR重建算法中的重要研究点。本文提出了一种融合M-估计和双边滤波的新型正则项框架,由于融合了M-估计对于边缘点的鲁棒性处理机制以及双边滤波的双重异性加权机制,因而相比于经典正则项具有更好的边缘保持特性。双边全变差在保持边缘的同时能够很好得抑制噪声,是目前使用较多的正则项,然而在噪声较大时它会出现“阶梯效应”,视觉效果差。研究发现在新型正则项框架下,选择Huber估计可以减轻“阶梯效应”,进一步提升算法的抗噪能力。合成序列和真实序列的实验结果都表明,这种基于Huber估计的新型正则项具有较好的边缘保持特性,且在噪声较大时能有效地减轻“阶梯效应”。更多还原

【Abstract】 Super Resolution Reconstruction is becoming a research hotspot in image processing area. It refers to methods that utilize information from multiple low resolution observed images to achieve restoration at resolution higher than that of original data. It is a kind of information fusion technologies that sacrifices time band for spatial resolution.The fundamental of super resolution reconstruction is presented along with degradation process modeling and regularized super resolution reconstruction algorithm. A detailed introduction to image quality assessment is also provided. The robust performance of an algorithm is very important for practical use. Existing algorithm based on least square estimator is examined to understand the limitations of robust characteristics. A cost function using M-estimators in regularized framework, which has better robust performance, is introduced. Then a comparison of common M-estimators is presented from two aspects, Huber estimator is proposed to use in cost function for its robust characteristic, estimation with small bias and convexity. And median absolute deviation is used for computing scale parameters of Huber estimator, which avoids error caused by artificial selection. Experiment results demonstrate that the proposed method has better performance than existing algorithms.Regularized term plays an important role in accelerating the convergence and stabilizing solution. It also affects visual effect of processing images, such as edge-preserving and so on. The design of regularized term for better visual perception is also a research hotspot in SR reconstruction. Based on bilateral filter and M-estimation, a novel regularized term is proposed. Because of combination of robustness of M-estimation and double weighting idea of bilateral filter, hence behaves much better in edge-preserving. Bilateral Total Variation is widely used in existing algorithms for its good noise-suppressing and edge-preserving ability, but it has“staircase effect”when suffering from severe noise. Then selection of M-estimator is examined for reducing“staircase effect”. Several experiments are designed to verify the effectiveness of this new regularized term using Huber estimator.更多还原