【作者】 王卓峥；

【导师】 贾克斌；

【作者基本信息】 北京工业大学 ， 电路与系统， 2013， 博士

【摘要】 随着计算机技术与信息处理技术的飞速发展，人们越来越发现利用现有硬件技术，提高信息获取质量的重要性。如何通过软件的方式突破图像固有信息的限制，提高图像分辨率，从而获得更多图像的高频细节信息，具有重要的研究意义与应用价值。在上述背景下，本文针对现有图像超分辨率重建方法在分辨率提升倍数较大时，算法复杂度高，重建效果不稳定的问题，研究并实现可针对通用图像的高性能超分辨率图像重建方法，主要研究成果包括：提出了基于主元分析与高斯加权欧式距离的快速图像配准方法PCA-SIFT-Gaussian。首先分析了特征提取与特征匹配的常用方法及具有竞争力的传统SIFT方法的四个主要步骤：图像尺度空间的建立、精确定位极值点、关键点方向分配和SIFT特征描述子的生成，使提取的特征对图像间发生平移、旋转、仿射变换、视角变换、光照变换等具有较好的不变性。针对传统SIFT算法中128维特征向量占用存储空间大，匹配耗时多等问题，围绕关键点的41×41的邻域内形成3042维特征向量，引入主元分析法对多维数据进行降维，完成特征提取。在特征匹配过程中采用高斯加权欧式距离代替传统的欧式距离进行阈值判定。实验结果证明，本文算法在保持较高性能的同时，特征提取速度相比传统SIFT方法提升约40%。对高斯噪声、旋转与尺度变化、仿射变换及光照变化等具有更高的鲁棒性。提出具有更快词典训练速度与重建精度的基于小波多分辨率分析的超分辨率图像重建方法。首先分析了稀疏表示法中三个重要步骤：建立图像降质模型、局部约束块稀疏表示、全局约束稀疏表示。针对稀疏表示法中训练过完备联合词典对计算量较大，算法耗时的问题，提出采用提升小波变换分解词典素材为个低频分量和三个高频分量，通过第二层和第三层提升小波变换后的高频分量估计出第层的高频分量，从而节约75%的像素数，降低训练词典的时间。在局部先验约束中特征提取算子F的选择上，采用PCA-SIFT-Gaussian代替传统维梯度高频滤波器以获得更多的高频细节。实验结果证明，本文算法与传统方法比较可缩短词典训练时间60%以上，且具有更高的重建精度。提出同时具备填补矩阵丢失元素与修复受损元素功能的新方法。首先通过比较不同矩阵填充与矩阵恢复主流方法，引入具有优越的收敛性及求解精度的增广拉格朗日乘子法，提出同时具备矩阵填充与矩阵恢复功能的新方法——双非精确增广拉格朗日乘子法（Dual-IALM）。实验结果表明，该方法具有迭代次数少、速度快、精度高的特点及较强的抗噪能力，通过该方法可较好地解决图像去噪与图像融合问题等实际应用问题。实现了由个或多个摄像头捕捉的具有亚像素位移的连续多帧低分辨率图像序列的高性能图像超分辨率重建系统。首先通过多帧图像配准方法PCA-SIFT-Gaussian对齐LR观测图像，然后采用第三章提出的改进的稀疏表示法训练联合词典对，并建立块稀疏表示和全局约束稀疏表示，生成初步重建的HR图像序列。最后采用Dual-IALM对图像序列构造的观测矩阵进行矩阵填充与矩阵恢复，将输出的低秩矩阵按照光栅扫描顺序重构为最终的HR图像。为了进步优化软件系统，预先保存配准后的特征向量，直接在特征提取算子F中使用。此外，本文引入Ring-Jacobi排序法取代传统的Round-Robin排序法减少并行算法中奇异值分解的迭代次数。实验结果证明，本文提出的新方法SRMCR相比当前其它主流超分辨率重建方法，在分辨率提升四倍以上时，仍保留丰富的图像细节，重建HR图像的峰值信噪比平均值高于其它方法5.04dB至6.28dB，且不易受词典训练素材选择范围的影响，能有效处理针对通用图像的高性能超分辨率重建问题，可应用于遥感图像超分辨率重建等机器视觉领域。更多还原

【Abstract】 With the rapid development of computer technology and information processingtechnology, people are increasingly found the importance of improving the quality ofthe information by using existing hardware technology. How to enhance the imageresolution through limit of the inherent information of the image, thereby obtainingmore high-frequency image details, which has important significance researching andapplication value.Based on the abovementioned background, to solve the complexity of thecomputation and unstable reconstruction effect in traditional super resolution methods,one high performance super resolution general-purpose images reconstruction methodis achieved in this paper in the following aspects:A fast image registration method, namely PCA-SIFT-Gaussian is proposed basedon Principal Component Analysis and Gaussian weighted Euclidean distance. Firstly,we introduced the most popular methods of feature extraction and matching. There arefour main steps to extract features by SIFT, Establishment of scale space, detection ofextrema, orientation assignment and generation of SIFT descriptor. To solve the largestorage space and matching time-consuming issues, we introduced3042dimensionalfeature vector surrounding the41×41pixels’ neighborhood. Then, PCA is achieved toreduce the multi-dimensional data instead of old128dimensional feature vectors. Inthe process of feature matching, we used Gaussian weighted Euclidean to replacetraditional Euclidean distance, which will meet the human visual characteristics better.Experimental results show that the proposed algorithm is more effective with fasterapproximately40%speed of feature extraction and matching compared to othermainstream image registration methods. It has higher robustness to Gaussian noise,rotation, and scale changes, affine transformation and illumination variation.An effective super resolution image reconstruction method with faster speed oftraining joint dictionary pair is proposed. This method is based on Lifting WaveletTransform and sparse representation. According to the three main steps of sparse representation, which are image degradation model, local constraints with patchessparse representation, and global constraints, we improved the process of trainingjoint dictionary pair. High-frequency components of the image can be estimated bycalculating only25%pixels in the whole image. That means this presented methodsave up to75%time to train dictionary. On the other hand, PCA-SIFT-Gaussianpresented before is used to achieve the feature extraction factor F instead oftraditional one-dimensional gradient high-pass filter. Experimental results show thatthe proposed algorithm can effectively shorten the dictionary training time by morethan60%, and has higher reconstruction accuracy.A new compressed sensing method, namely Dual-IALM, is proposed to achievematrix completion and matrix recovery simultaneously. After comparing allmainstream methods of matrix completion and matrix recovery, we introduced theAugmented Lagrangian Multiplier method, which has superior convergence andaccuracy and present a new method with this two matrix functions——Dual InexactAugmented Lagrangian Multiplier, abbreviated as Dual-IALM. Experimental resultsshow that this method has fewer iterations times, higher accuracy and strongeranti-noise ability, this method can solve the practical application of image denoisingand image fusion effectively.A high performance image super resolution reconstruction system is achieved formulti-frame sub-pixel displacement low-resolution image sequences captured by oneor more sensors. Super-resolved reconstruction of images can yield poor results in theabsence of extensively-trained related dictionaries. A super-resolution algorithm ispresented which remedies this problem by exploiting recent results from the work onsparse representation and matrix completion. An over-complete dictionary pair istrained using natural image data. Sparse coefficients of low-resolution image patchesare estimated using local prior constraints. In multi-frame images, sparse coefficientsare similar across frames, and the Inexact Augmented Lagrange Multiplier method isemployed to achieve matrix completion and recovery in the process of imposingglobal constraints. Furthermore, we optimize the system with keeping thePCA-SIFT-Gaussian descriptor and introducing Ring-Jacobi ordering to accelerate the singular value decomposition. The final high-resolution image is generated from theoutput low-rank matrix. Experiments reveal that the method yields higher PSNR value(from5.04dB to6.28dB)than other mainstream SR algorithms, produces perceptiblysuperior edges and details, and is more robust to dictionary insufficiency which can becan be applied to the field of remote sensing and machine vision.更多还原