【作者】 王艳；

【导师】 王新年；

【作者基本信息】 大连海事大学 ， 信息与通信工程， 2012， 硕士

【摘要】 图像超分辨率重建的目的是根据单幅或者多幅低分辨率图像,采用信号处理技术,重建出高分辨率图像,其可分为单幅图像超分辨率重建和多幅图像超分辨率重建,本文主要研究单幅图像超分辨率重建。基于稀疏表示的超分辨率重建技术是当前图像处理领域研究的热点之一,其关键技术包括稀疏矩阵的计算、词典构建以及相关后处理技术。本文在分析词典构建算法以及训练样本集对重建质量影响分析的基础上,提出了改进的基于稀疏表示的彩色图像超分辨率重建方法,并对图像后处理技术进行了研究。主要工作包括：1)词典构建算法性能验证。实现了两种典型的词典构建算法,并进行了性能验证实验。实验结果表明KSVD算法在恢复相似度和平均表示误差这两个指标上均优于MOD算法。2)训练样本集及词典构建算法对超分辨率重建性能的影响分析。词典的构建需要样本集的训练,训练集一般分为两类：(1)自然图像训练集；(2)与待重建图像相关的训练集。将这两种训练集分别利用KSVD和MOD算法构建两类词典,并进行超分辨率重建。实验结果表明基于KSVD算法超分辨率重建的信噪比和相似度都高于MOD算法；训练集与待重建图像越相近,信噪比和相似度就越高。3)改进的基于稀疏表示的彩色图像超分辨率重建方法。将低分辨率图像的RGB模式转换成YCbCr模式,利用KSVD算法构建Y、Cb、Cr三通道词典,分别对三个通道进行图像的超分辨率重建。实验结果表明与单独的Y分量重建以及R、G、B三通道重建方法相比较,本文算法重建的信噪比和相似度都有了提高。4)基于几何局部自适应性锐化(GLAS)的图像后处理。针对重建后图像边缘模糊现象,采用GLAS算法进行图像后处理,根据图像的局部轮廓构建不同形状的核函数进行各向异性图像增强,实验取得较好效果。更多还原

【Abstract】 The purpose of image super-resolution reconstruction is to reconstruct a high-resolution image or sequence from a single image or multiple low-resolution images which is based on signal processing techniques.lt can be classified into single image super-resolution reconstruction and multiple images super-resolution reconstruction. This thesis focus on single image super-resolution reconstruction algorithms.Sparse representation based single super-resolution reconstruction technique is one of the hot topics in the field of image processing.The key technologies include sparse matrix computing and dictionary construction as well as related post-processing techniques. Based on the analysis of dictionary construction algorithms’performance and the effect of the training sample sets on reconstruction quality,an improved color image super-resolution reconstruction method based on sparse representation is proposed,and a post-processing algorithm is used to enhance reconstructed images.The main work are as follows:1) Performance verification of the dictionary construction algorithms.Two typical dictionary construction algorithms are implemented and the verification tests are conducted.The experimental results show that the KSVD algorithm outperforms MOD algorithm in respect of the relative number of correctly recovered atoms and the average representation error.2) Influence analysis of training sample sets and dictionary construction algorithms on the performance of super-resolution reconstruction.The construction of dictionary needs training sample sets, the training sets generally include two categories:(1) natural images training set;(2) the training set related to the image which is to be reconstructed. Two types of dictionaries are constructed respectively from two training sets with KSVD and MOD algorithms, and super-resolution reconstruction tests are conducted based on the two trained dictionaries. The experiment results show that reconstructed images based on the dictionary trained by the KSVD algorithm are better than the one trained by the MOD algorithm in respect of the peak signal to noise ratio(PSNR) and structural similarity (SSIM),and the more similar training set to the image which is to be reconstructed, the higher PSNR and SSIM are. 3) Proposing an improved method of color image super-resolution reconstruction algorithm based on sparse representation.The image to be reconstructed is firstly converted the RGB mode to the YCbCr mode,then the Y, Cb and Cr channel dictionaries are trained with the KSVD algorithm,and finally super-resolution reconstruction are performed on the three channels. The experiment results show that, compared with the separate Y component of the reconstruction algorithm and the R, G and B channel reconstruction algorithm, the proposed algorithm has better performance on PSNR and SSIM.4) Using a geometric locally adaptive sharpening (GLAS) based image post-processing algorithm to enhance the reconstructed image. For the reason that the reconstructed image edge is always blurred, a GLAS algorithm is used to enhance the image edges. The kernel functions are constructed according to the different shapes and sizes of the local structure of the image,and then are used to perform anisotropic filtering on the input image.The experimental results show that post-processing is necessary and can improve the reconstructed image quality numerically and visually.更多还原