Print

发布时间: 2018-09-16
摘要点击次数:
全文下载次数:
DOI: 10.11834/jig.180011
2018 | Volume 23 | Number 9




    图像处理和编码    




  <<上一篇 




  下一篇>> 





核稀疏表示和原子相关度的图像重建
expand article info 程德强, 刘威龙, 邵丽蓉, 陈亮亮
中国矿业大学信息与控制工程学院, 徐州 221008

摘要

目的 针对目前基于稀疏表示的超分辨率重建算法中对字典原子的选取效率低、图像重建效果欠佳的问题,本文提出了核方法与一种高效的字典原子相关度筛选方法相融合的图像超分辨重建算法,充分利用字典原子与图像的相关度,选用对重建的贡献最大的原子来提高重建的效率和效果。方法 首先,通过预处理高分辨率图像得到高、低分辨率图像样本集,再用字典学习得到高、低分辨率字典对;然后,对字典原子进行非相关处理提高字典原子的表达能力;此后,再利用低分辨率字典,引入核方法和字典原子筛选方法进行稀疏表示,设置阈值筛选高相关原子,低相关度原子对重建贡献度低,在迭代过程中耗费计算量,所以舍去低相关原子,再对普通原子进行正则化处理后加入支撑集,处理后的字典原子对于重建具有很好的表达能力;最后,利用处理后的字典原子对低分辨率图求解稀疏表示问题得到稀疏表示系数,结合高分辨率字典重建出高分辨率图像。结果 实验通过与其他学习算法对比,得到结构相似度(SSIM)、峰值信噪比(PSNR)以及重建时间的结果。实验结果表明:本文方法与对比方法相比,图像重建时间提高了22.2%;图像结构相似度提高了9.06%;峰值信噪比提高了2.30 dB。原有的基于字典学习的方法对于字典选取具有一定的盲目性,所选取的原子与重建图像相关度较低,使重建效果差,本文方法获得的字典原子可以减少稀疏表示过程的时耗,同时提高稀疏表示的精度。引入核方法,改善经典算法中对原子选取的低精度问题,经实验证明,本方法能有效提高重建算法性能。结论 实验结果表明,图像的稀疏表示过程的重建时间明显减少,重建效果也有一定的提高,并且在训练样本较少的情况下同样有良好的重建效率和效果,适合在实际中使用。

关键词

稀疏表示; 超分辨率重建; 核方法; 原子相关度; 非相关处理

Super resolution reconstruction algorithm based on kernel sparse representation and atomic correlation
expand article info Cheng Deqiang, Liu Weilong, Shao Lirong, Chen Liangliang
School of Information and Control Engineering, China University of Mining and Technology, Xuzhou 221008, China
Supported by: National Natural Science Foundation of China(51774281)

Abstract

Objective To overcome the low efficiency of dictionary atom screening and the poor effect of image reconstruction results in some super-resolution methods based on sparse representation, which are mostly unconsidered in atom screening, this paper proposes a super-resolution reconstruction algorithm. This algorithm is based on a combination of kernel method and dictionary atomic correlation, which fully uses the correlation between the dictionary and image, and selects the atoms, which significantly contributes to the reconstruction results and improves the efficiency and effect of the reconstruction. Method First, a set of low-and high-resolution samples are obtained by pre-processing applied on the high-resolution images. Low-and high-resolution dictionaries are learned by using a dictionary learning algorithm. Second, the dictionary atom is uncorrelated to improve the ability of the dictionary atom to express. Third, by using the low-resolution dictionary, the kernel method and dictionary atom screening method are used for sparse representation, to set thresholds to screen for highly correlated atoms, eliminate low-correlation atoms, and then use the normal atoms for normalized processing. The resulting high-and low-resolution dictionary atoms are incoherent, thereby eliminating the similarity between dictionary atoms, enhancing the expressive power of dictionary atoms, and helping to select the next dictionary atoms. In the process of solving the representation coefficient, selecting the appropriate atoms from the low-resolution dictionary to the support set, which is the largest part of the computation, is necessary. When updating the support set, the dictionary of low-resolution images is trained from other images, which leads to the large contribution of some atoms to the samples. The atoms with low correlation often do not contribute during the iteration process, but each iteration has considerable computation costs. At the same time, for the image blocks that need to be restored, a number of highly correlated atomic pairs have a major contribution to reconstruction. To reduce the computational complexity and improve the reconstruction effect, this paper improves the traditional method by using the correlation of the residual and atom to conduct efficient dictionary selection. Finally, the sparse representation problem is solved to obtain sparse coefficients, and the super-resolution image is recovered by using these coefficients. High-resolution image blocks are obtained by using high-resolution dictionary and coefficient representation coefficients, and then, high-resolution images are synthesized by utilizing image blocks. Result The performance of algorithm reconstruction is measured by PSNR, structural similarity, and time and compared with Yang, MSDSC, and SDCKR algorithms. In the experiment, the following test chart is analyzed in detail, and the ImageNet standard image database is trained to obtain additional detailed experimental results. The experimental results show that, compared with the contrast method, the image reconstruction time is increased by 22.2%, the image structure similarity is increased by 9.06%, and the PSNR is increased by 2.30 dB. The original method based on dictionary learning for dictionary selection has a certain blindness. The atom and reconstruction image correlation degree is low, and the reconstruction effect is poor. This method can reduce the dictionary sparse representation of time consumption and improve the accuracy of sparse representation. In the super-resolution reconstruction of the classical image reconstruction algorithm, the effect is not ideal and the reconstruction time is too long. The main reason is that the dictionary selection efficiency is low, aiming at the abovementioned problem. For the improvement of dictionary learning algorithm in solving the sparse coefficient method in the process of nuclear innovation and the introduction of machine learning and new atom selection method, this paper presents a test with the commissioning of a large number of practical engineering images. The experimental results show that this method can improve the reconstruction effect and reduce the time required for reconstruction. Conclusion Compared with the same algorithm of dictionary learning, the reconstruction time of this algorithm is also less. Experiments have proven that in this method, the reconstruction time of image sparse representation process is significantly reduced. The reconstruction effect is also improved, with good reconstruction efficiency and effectiveness under the condition of few training samples, which is suitable for practical use.

Key words

sparse representation; super-resolution reconstruction; kernel method; atomic correlation; unrelated processing

0 引言

图像分辨率是图像质量的重要指标, 分辨率越高, 像素的密度越大,提供的信息就越丰富。在遥感监测、军事侦察、交通及安全监控、医学诊断和模式识别等应用中, 都需要高分辨率图像。由于受成像系统物理条件和环境的影响等, 在成像过程中常常存在光学和运动模糊、下采样和噪声等退化过程, 使实际得到的图像分辨率低。

超分辨率概念最早出现在光学领域。在这个领域中,超分辨率是指试图复原衍射极限以外数据的过程, 它是利用一幅低分辨率(LR)图像或者图像序列,重构出包含更多细节的高分辨率(HR)图像,它可以显著提高图像质量而不需要增加硬件成本,在视频监控、医学成像、遥感图像等领域中已获得了巨大成就。最初的重建算法是基于模型的方法,其特点是效率高,但重建效果欠佳,出现许多改进算法。Yu等人[1]在基本混合粒子群优化算法下改进,提高了重建的有效性;Zhao等人[2]基于混合非局部先验模型的算法,针对重建的病态特性,利用多框架的重建方法、贝叶斯先验模型等更好地保存了图像的边缘、纹理特性;Huang等人[3]的梯度矢量流混合场模型算法,在原来的多框架模型下加入图像增强与去噪,提高了最终重建的精度;Wei等人[4]的高阶导数插值联合分数滤波函数算法,在傅氏变换域使用高阶导数插值方法,提高高频信息的重建精度。而基于字典学习的方法可以从其他图像中获取先验信息,大大降低了对输入图像的要求,可以对单幅图像完成超分辨率重建[5],使得近年来基于学习的图像超分辨率重建算法引起了国内外众多研究者的关注。Yang等人[6-7]开创性地将稀疏表示理论引入到图像超分辨率重建,利用同一幅图的高、低分辨率图像块在特定的稀疏基下有相同的稀疏表示系数作为约束条件,训练得到高、低分辨率图像字典;Zhu等人[8]的多稀疏字典模型(MSDM)方法,分别对图像的垂直、水平、高频信息处理,从多个角度提高重建精度;Yang等人[9]将子字典与核回归方法引入提出核退化稀疏编码(SDCKR)算法,通过训练独立的高、低分辨率字典提高字典的表达能力,用局部子字典编码图像块,改善局部效果;Zeyde等人[10]在杨建超等人的基础上进行改进,利用PCA对训练样本的特征进行降维,并使用K-SVD(K次奇异值分解)的方法进行字典训练,提高了字典训练效率;Elad等人[11-13]利用冗余字典训练,引入机器学习的方法,在K-SVD、OMP(正交匹配追踪)等经典算法的基础上改进,使重建算法性能得到提高。

经典的图像超分辨率重建算法中,重建效果不够理想并且重构时间过长, 主要原因在于字典原子的选择效率低, 本文针对以上问题,对字典学习算法改进,在求解稀疏表示系数过程中创新地引入机器学习的核方法以及新的原子选取方法, 本文对大量实际工程图像进行检验与调试, 实验结果表明,本文方法能提高重建效果并减少重建所需时间。

1 相关工作

1.1 图像退化降质模型

在超分辨率重建算法的字典训练过程中,需要得到原始高分辨率图像的降质图,用于低分辨率字典的训练。通常采用的退化模型为

$ {\mathit{\boldsymbol{Y}}_k} = {\mathit{\boldsymbol{C}}_k}{\mathit{\boldsymbol{B}}_k}{\mathit{\boldsymbol{H}}_k}{\mathit{\boldsymbol{K}}_k} + {\mathit{\boldsymbol{N}}_k} $ (1)

式中,$\mathit{\boldsymbol{X}}$${\bf{{\bf{R}}}^\rm{H}}$表示HR图像, $Y$${{\bf{R}}^{\rm{L}}}$表示降噪LR图像, $k$表示图像或图像块编号, $\mathit{\boldsymbol{C}}$表示下采样矩阵, $\mathit{\boldsymbol{B}}$表示光学模糊矩阵, $\mathit{\boldsymbol{H}}$表示$\mathit{\boldsymbol{X}}$相对于$\mathit{\boldsymbol{Y}}$的运动变形矩阵, $\mathit{\boldsymbol{N}}$表示加性随机白噪声。

1.2 图像稀疏表示模型

基于稀疏表示理论,图像在过完备字典下,总存在稀疏的表示,即大部分系数为零,只有少数的非零系数。假设$\mathit{\boldsymbol{D}} \in {{\bf{R}}^{n \times K}}$是维度为$n$、原子数为$K$的过完备字典,$\mathit{\boldsymbol{X}} \in {{\bf{R}}^n}$为图像子块的向量表示,则$\mathit{\boldsymbol{X}}$的稀疏表示的模型[14]

$ \begin{array}{*{20}{c}} {\min {{\left\| \mathit{\boldsymbol{\alpha }} \right\|}_0},}&{{\rm{s}}.\;{\rm{t}}.\;\mathit{\boldsymbol{X}} = \mathit{\boldsymbol{Da}}} \end{array} $ (2)

式中,$\mathit{\boldsymbol{\alpha }} \in {\bf{R}}$$\mathit{\boldsymbol{X}}$在字典$\mathit{\boldsymbol{D}}$下的最稀疏表示,可通过正交匹配追踪算法等贪婪算法来近似求解。可将上述问题转换为凸优化的问题,从而式(2)可以转换为

$ \begin{array}{*{20}{c}} {\min {{\left\| \mathit{\boldsymbol{\alpha }} \right\|}_1},}&{{\rm{s}}.\;{\rm{t}}.\;\left\| {\mathit{\boldsymbol{Da}} - \mathit{\boldsymbol{X}}} \right\|_2^2 \le \varepsilon } \end{array} $ (3)

式中,$\varepsilon $为逼近精度。对于基于字典学习的图像超分辨率重建,需要两个过完备字典,其中一个是在LR特征空间下训练得到的LR字典,用于对LR图像进行稀疏分解,另一个是在HR特征空间下训练得到的HR字典,用于最终的HR图像重建[15],分别记作${\mathit{\boldsymbol{D}}_{\rm{L}}}$${\mathit{\boldsymbol{D}}_{\rm{H}}}$

对于HR图像的过完备字典和LR图像获得的过完备字典, 其稀疏模型$\mathit{\boldsymbol{X}}$=${\mathit{\boldsymbol{D}}_{\rm{H}}}$$\mathit{\boldsymbol{\alpha }}$$\mathit{\boldsymbol{X}}$=${\mathit{\boldsymbol{D}}_{\rm{L}}}$$\mathit{\boldsymbol{\alpha }}$,存在一个映射矩阵$\mathit{\boldsymbol{H}}$使得${\mathit{\boldsymbol{D}}_{\rm{L}}}$=$\mathit{\boldsymbol{H}}$${\mathit{\boldsymbol{D}}_{\rm{H}}}$, 即高低分辨率图像块在对应过完备字典下具有相同最稀疏表示,因此,当建立了合适的字典对后,由$\mathit{\boldsymbol{Y}}$=${\mathit{\boldsymbol{D}}_{\rm{L}}}$$\mathit{\boldsymbol{\alpha }}$得到对应的稀疏系数$\mathit{\boldsymbol{\alpha }}$,结合HR图像得到的过完备字典${\mathit{\boldsymbol{D}}_{\rm{H}}}$,利用$\mathit{\boldsymbol{X}}$=${\mathit{\boldsymbol{D}}_{\rm{H}}}$$\mathit{\boldsymbol{\alpha }}$就可以恢复出高分辨率图像块。

2 图像超分辨率重建算法

2.1 稀疏字典对训练

求解高、低分辨率字典对的目标函数可以表示为

$ \begin{array}{*{20}{c}} {\left\{ {\mathit{\boldsymbol{D}},\mathit{\boldsymbol{A}}} \right\} = \mathop {\arg \min }\limits_{\mathit{\boldsymbol{D}},\mathit{\boldsymbol{A}}} \left\| {\mathit{\boldsymbol{X}} - \mathit{\boldsymbol{DA}}} \right\|_{\rm{F}}^2}\\ {{\rm{s}}.\;{\rm{t}}.\;\;\forall i = 1,2, \cdots ,k,{{\left\| {{\mathit{\boldsymbol{\alpha }}^i}} \right\|}_0} \le T} \end{array} $ (4)

式中,$\mathit{\boldsymbol{A = }}{\left[{{\mathit{\boldsymbol{\alpha }}^1}, {\mathit{\boldsymbol{\alpha }}^2}, \cdots, {\mathit{\boldsymbol{\alpha }}^i}} \right]^{\rm{T}}}$表示系数矩阵,$T$是系数的稀疏约束项。

训练样本是由HR图像和经过降质处理的LR图像的联合样本集合$\mathit{\boldsymbol{S}} = \{ \mathit{\boldsymbol{X}}, \mathit{\boldsymbol{Y}}\} $,其中,$\mathit{\boldsymbol{X}} = \{ {\mathit{\boldsymbol{X}}_1}, {\mathit{\boldsymbol{X}}_2}, \cdots, {\mathit{\boldsymbol{X}}_n}\} $, $\mathit{\boldsymbol{Y}} = \{ {\mathit{\boldsymbol{Y}}_1}, {\mathit{\boldsymbol{Y}}_2}, \cdots, {\mathit{\boldsymbol{Y}}_n}\} $,训练过程可以表示为

$ \left\{ {{\mathit{\boldsymbol{D}}_{\rm{H}}},\mathit{\boldsymbol{A}}} \right\} = \mathop {\arg \min }\limits_{{\mathit{\boldsymbol{D}}_{\rm{H}}},\mathit{\boldsymbol{A}}} \left\| {\mathit{\boldsymbol{X}} - {\mathit{\boldsymbol{D}}_{\rm{H}}}\mathit{\boldsymbol{A}}} \right\|_{\rm{F}}^2 $ (5)

$ \left\{ {{\mathit{\boldsymbol{D}}_{\rm{L}}},\mathit{\boldsymbol{A}}} \right\} = \mathop {\arg \min }\limits_{{\mathit{\boldsymbol{D}}_{\rm{L}}},\mathit{\boldsymbol{A}}} \left\| {\mathit{\boldsymbol{X}} - {\mathit{\boldsymbol{D}}_{\rm{L}}}\mathit{\boldsymbol{A}}} \right\|_{\rm{F}}^2 $ (6)

高、低分辨率字典具有相同的稀疏因子,所以可将两个字典的训练合并在一个编码框架下[7],即

$ \frac{1}{M}\left\{ {{\mathit{\boldsymbol{D}}_{\rm{H}}},\mathit{\boldsymbol{A}}} \right\} + \frac{1}{N}\left\{ {{\mathit{\boldsymbol{D}}_{\rm{L}}},\mathit{\boldsymbol{A}}} \right\} = \min \left\| {\mathit{\boldsymbol{Z}} - \mathit{\boldsymbol{DA}}} \right\|_{\rm{F}}^2 $ (7)

$ \left\{ \begin{array}{l} \mathit{\boldsymbol{Z}} = {\sqrt M ^{ - 1}}\mathit{\boldsymbol{X/}}{\sqrt N ^{ - 1}}\mathit{\boldsymbol{Y}}\\ \mathit{\boldsymbol{D}} = {\sqrt M ^{ - 1}}{\mathit{\boldsymbol{D}}_{\rm{H}}}\mathit{\boldsymbol{/}}{\sqrt N ^{ - 1}}{\mathit{\boldsymbol{D}}_{\rm{L}}} \end{array} \right. $ (8)

$M$$N$分别为样本中图像块向量形式的维度。可以用$\mathit{\boldsymbol{Z}}$作为K-SVD[16]算法的输入,获得字典$\mathit{\boldsymbol{D}}$,再用式(8)求得高、低分辨率字典。

为了提高字典的原子间的非相干性,本文引入基于梯度的方法[17],在字典训练的过程中对字典原子处理,设${\mathit{\boldsymbol{D}}_i}$是第$i$次迭代的字典,即

$ {\mathit{\boldsymbol{D}}_i} = \mathop {\arg \min }\limits_{{\mathit{\boldsymbol{D}}_i}} \left\| {\mathit{\boldsymbol{D}}_i^{\rm{T}}{\mathit{\boldsymbol{D}}_i} - \mathit{\boldsymbol{I}}} \right\|_{\rm{F}}^2 $ (9)

式中, $\mathit{\boldsymbol{I}}$是一个单位矩阵,梯度$\mathit{\boldsymbol{E}}{\bf{:}} = \left\| {\mathit{\boldsymbol{D}}_i^{\rm{T}}{\mathit{\boldsymbol{D}}_i}-\mathit{\boldsymbol{I}}} \right\|_{\rm{F}}^2$,插入到$\mathit{\boldsymbol{D}} \leftarrow \mathit{\boldsymbol{D}}-\eta \nabla \mathit{\boldsymbol{E}}$

$ \mathit{\boldsymbol{D}}_i^{{\rm{new}}} = \mathit{\boldsymbol{D}}_i^{{\rm{old}}} - \eta \mathit{\boldsymbol{D}}_i^{{\rm{old}}}\left( {\mathit{\boldsymbol{D}}_i^{{\rm{Told}}}\mathit{\boldsymbol{D}}_i^{{\rm{old}}} - \mathit{\boldsymbol{I}}} \right) $ (10)

由此得到的高、低分辨率字典原子具有不相干性,消除了字典原子间的相似性,增强了字典原子的表达能力,有利于接下来字典原子的选取。

2.2 稀疏表示系数求解

在求解表示系数的过程中,需从低分辨率字典中选取合适的原子加入到支撑集,这是计算量最大的部分。在更新支撑集的时候,由于低分辨率图像的字典是从其他图像训练得到的,容易使得某些原子对样本的表示贡献度非常低。当相关度为零时,该原子甚至会导致字典无法更新,相关性极低的原子在迭代过程中往往无贡献但每次迭代耗费计算量。同时对于需要恢复的图像块,往往有若干相关性极高的原子对重建有主要的贡献。为了减小计算复杂度,同时改善重建效果,本文改进了传统方法,利用残差与原子的相关性进行高效的字典选取[18]

本文引入机器学习中的核方法,将所需处理的数据通过非线性映射$\mathit{\Phi }$映射到高维(甚至无穷维)空间$\mathfrak{R}$中,$\mathfrak{R}$是一个希尔伯特空间,则非线性映射$\mathit{\Phi }$对应的核函数表示为

$ \begin{array}{*{20}{c}} {\left( {\mathit{\Phi }\left( \mathit{\boldsymbol{x}} \right),\mathit{\Phi }\left( {\mathit{\boldsymbol{x'}}} \right)} \right) = }\\ {\mathit{\Phi }{{\left( \mathit{\boldsymbol{x}} \right)}^{\rm{T}}}\mathit{\Phi }\left( {\mathit{\boldsymbol{x'}}} \right) = k\left( {\mathit{\boldsymbol{x}},\mathit{\boldsymbol{x'}}} \right)} \end{array} $ (11)

式中,$k$为核函数,其求解不需知道具体的映射$\mathit{\Phi }$,并且与数据的维度无关,只需选取合适的核函数$k$,可以减小向量内积的计算。

同时引入相似度公式,设,$\mathit{\boldsymbol{X}} = \{ {\mathit{\boldsymbol{X}}_1}, {\mathit{\boldsymbol{X}}_2}, \cdots, {\mathit{\boldsymbol{X}}_n}\} $$\mathit{\boldsymbol{Y}} = \{ {\mathit{\boldsymbol{Y}}_1}, {\mathit{\boldsymbol{Y}}_2}, \cdots, {\mathit{\boldsymbol{Y}}_n}\} $,其相似度为

$ \rho = \mathit{\boldsymbol{XY/}}\sqrt {{\mathit{\boldsymbol{X}}^2}} \sqrt {{\mathit{\boldsymbol{Y}}^2}} $ (12)

现在对需要处理的图像$\mathit{\boldsymbol{Y}}$进行稀疏分解。首先通过插值法将图像$\mathit{\boldsymbol{Y}}$恢复到高分辨图像相同尺寸的低分辨插值图像${\mathit{\boldsymbol{Y}}_{\rm{L}}}$,并分割成$N$×$N$的图像块并将图像块向量化处理,${\mathit{\boldsymbol{Y}}_{\rm{L}}} = \left( {\mathit{\boldsymbol{y}}_{\rm{L}}^1, \mathit{\boldsymbol{y}}_{\rm{L}}^2, \cdots, \mathit{\boldsymbol{y}}_{\rm{L}}^k, \cdots, \mathit{\boldsymbol{y}}_{\rm{L}}^n} \right)$,其中${\mathit{\boldsymbol{y}}_{\rm{L}}^k}$${\mathit{\boldsymbol{Y}}_{\rm{L}}}$的第$k$个图像块。求解各个图像块的稀疏表示系数

$ \begin{array}{*{20}{c}} {{\mathit{\boldsymbol{\alpha }}_k} = \mathop {\arg \min }\limits_{{\mathit{\boldsymbol{\alpha }}_k}} \left\| {\mathit{\boldsymbol{y}}_{\rm{L}}^k - {\mathit{\boldsymbol{D}}_\Lambda }} \right\|_2^2}\\ {{\rm{s}}.\;{\rm{t}}.\;\;{{\left\| {{\mathit{\boldsymbol{\alpha }}_k}} \right\|}_0} \le {T_0}} \end{array} $ (13)

式中,${T_0}$是稀疏约束项,$\mathit{\Lambda }$是支撑集,${\mathit{\boldsymbol{D}}_\Lambda }$是由支撑集对应${\mathit{\boldsymbol{D}}_{\rm{L}}}$中的原子构成的字典。式(13)求解的方法借鉴OMP算法:先固定系数${\mathit{\boldsymbol{\alpha }}_k}$,更新支撑集$\Lambda $。残差$r$的公式为

$ \mathit{\boldsymbol{r}} = \mathit{\boldsymbol{y}}_{\rm{L}}^k - {\mathit{\boldsymbol{D}}_\Lambda }{\mathit{\boldsymbol{\alpha }}_k} $ (14)

本文需要残差$\mathit{\boldsymbol{r}}$与字典原子${\mathit{\boldsymbol{d}}_{{{\rm{L}}_{\rm{i}}}}}$的相关系数,引入核函数,与${\mathit{\boldsymbol{d}}_{{{\rm{L}}_{\rm{i}}}}}$映射到核空间可得到$\mathit{\Phi }\left( \mathit{\boldsymbol{r}} \right)$$\mathit{\Phi }\left( {{\mathit{\boldsymbol{d}}_{{{\rm{L}}_{\rm{i}}}}}} \right)$代入式(11)(12),得

$ \begin{array}{*{20}{c}} {{\rho _i} = \mathit{\Phi }\left( \mathit{\boldsymbol{r}} \right)\mathit{\Phi }\left( {{\mathit{\boldsymbol{d}}_{{\rm{Li}}}}} \right)/\sqrt {{\mathit{\Phi }^2}\left( \mathit{\boldsymbol{r}} \right)} \sqrt {{\mathit{\Phi }^2}\left( {{\mathit{\boldsymbol{d}}_{{\rm{Li}}}}} \right)} = }\\ {k\left( {\mathit{\boldsymbol{r}},{\mathit{\boldsymbol{d}}_{{\rm{Li}}}}} \right)/\sqrt {k\left( {\mathit{\boldsymbol{r}},\mathit{\boldsymbol{r}}} \right)} \sqrt {k\left( {{\mathit{\boldsymbol{d}}_{{\rm{Li}}}},{\mathit{\boldsymbol{d}}_{{\rm{Li}}}}} \right)} } \end{array} $ (15)

代入核函数公式求出残差与第$i$个原子的相关系数${\rho _i}$,本文核函数取高斯核函数:$k\left( {\mathit{\boldsymbol{x}}, \mathit{\boldsymbol{x'}}} \right) = {\rm{exp}}\left( {-\left\| {\mathit{\boldsymbol{x}}-\mathit{\boldsymbol{x'}}} \right\|/2{\delta ^2}} \right)$

根据相关系数对原子进行处理:若相关系数小于阈值${\rho _a}$,这些原子对于此图像块重建的关联低,每次都迭代会耗费计算,从字典中除去此类贡献低的原子;若相关系数大于阈值${\rho _b}$,对于这些对重建贡献高的原子进行正则化处理,加入支撑集;若相关系数${\rho _i}$∈[${\rho _a}$, ${\rho _b}$],只取最大的相关系数对应的原子加入支撑集,由于只有一个原子,不需要正则化处理。本文${\rho _a}$取0.2,${\rho _b}$取0.8, 在实验结果分析中会具体分析阈值的取值问题。

新的原子确定后,更新支撑集并固定支撑集$\Lambda $,则${\mathit{\boldsymbol{\alpha }}_k}$可以由式(13)求出。迭代直至收敛得到最终的表示系数${\mathit{\boldsymbol{\alpha }}_k}$。算法步骤如下:

输入:低分辨率图像块$\mathit{\boldsymbol{y}}_{\rm{L}}^k$,低分辨率字典${\mathit{\boldsymbol{D}}_{\rm{L}}}$

初始化:支撑集$\mathit{\boldsymbol{ \boldsymbol{\varLambda} }}$为空集,稀疏表示系数${\mathit{\boldsymbol{\alpha }}_k}$=0。

1) 利用式(14)求出残差$\mathit{\boldsymbol{r}}$,用式(15)求解相关系数${\rho _i}$

2) if ${\rho _i}$>${\rho _a}$

then do正则化处理相应原子加入支撑集

else if ${\rho _a}$${\rho _i}$${\rho _b}$

then do $\mathit{\boldsymbol{ \boldsymbol{\varLambda} }}$$\mathit{\boldsymbol{ \boldsymbol{\varLambda} }}$$\mathop {{\rm{arg}}\;{\rm{max}}}\limits_i \;{\rho _i}$

else do除去对应原子

end if

3) 再利用式(13)更新稀疏表示系数${\mathit{\boldsymbol{\alpha }}_k}$

4) 重复以上步骤,直至收敛。

输出:稀疏表示系数${\mathit{\boldsymbol{\alpha }}_k}$

2.3 重建

利用高分辨率字典和系数表示系数得到第$k$个高分辨率图像块${x_k} = {\mathit{\boldsymbol{D}}_{\rm{H}}}{\mathit{\boldsymbol{\alpha }}_k}$,再由图像块合成高分辨率图像$\mathit{\boldsymbol{X}}$,即

$ \mathit{\boldsymbol{X}} = {\mathit{\boldsymbol{Y}}_{\rm{L}}} + {\left( {\sum\limits_k {\mathit{\boldsymbol{R}}_k^{\rm{T}}{\mathit{\boldsymbol{R}}_k}} } \right)^{ - 1}}\left( {\sum\limits_k {\mathit{\boldsymbol{R}}_k^{\rm{T}}{\mathit{\boldsymbol{R}}_k}} } \right) $ (16)

式中,${\mathit{\boldsymbol{R}}_k}$是用于提取第$k$个图像块的矩阵。

3 实验分析

实验在CPU为双核1.80 GHz,内存为8.00 GB的计算机上运行,使用软件为Matlab2012b,图像放大倍数为2,图像块尺寸为17×17像素,算法重建性能的衡量采用峰值信噪比(PSNR)、结构相似度(SSIM)和时间,并与Yang[6]的算法、MSDSC[8]算法和SDCKR[9]算法对比。实验对测试图做了详细分析,并用Imagenet标准图像数据库训练,得到更详细实验结果(图 1)。

图 1 测试图重建结果
Fig. 1 Reconstruction of test images
((a)input images; (b)result of Yang; (c)result of MSDSC; (d)result of SDCKR; (e) ours)

相比于Yang算法、MSDSC算法、SDCKR算法,本文算法重建的图像,从主观来看图像的细节更加清晰;通过分析表 1表 2所示的客观评价指标,本文算法重建图像的PSNR的均值提高了2.30 dB,SSIM的均值提高了0.075,重建时间相对于字典学习算法平均提高了66.5 s。因此,理论分析和实验结果表明,本文算法对提高图像重建质量和重建速度有良好效果。

表 1 测试图像重建的PSNR、SSIM和时间对比
Table 1 Comparison of test images reconstruction on PSNR, SSIM and time

下载CSV
图像 方法 PSNR/dB SSIM 时间/s
Lena Yang 33.78 0.901 0 289.0
MSDSC 34.71 0.954 6 202.5
SDCKR 32.32 0.865 1 238.2
本文 36.65 0.989 3 129.2
Bike Yang 34.45 0.867 0 198.8
MSDSC 35.51 0.932 3 280.1
SDCKR 29.12 0.875 1 170.2
本文 36.12 0.977 4 231.1
Girl Yang 34.77 0.816 5 265.9
MSDSC 35.54 0.860 4 370.1
SDCKR 34.81 0.894 9 347.9
本文 36.81 0.954 9 267.9
Grope Yang 26.87 0.855 4 399.3
MSDSC 27.14 0.897 1 339.5
SDCKR 29.89 0.822 8 430.0
本文 28.89 0.912 8 340.0

表 2 Imagenet Set5数据集图像重建的平均PSNR、SSIM和时间值对比
Table 2 Comparison of Set 5 images reconstruction on average PSNR, SSIM and time

下载CSV
方法 PSNR/dB SSIM 时间/s
Yang 25.22 0.765 5 321.6
MSDSC 26.56 0.867 0 265.2
SDCKR 25.67 0.849 2 310.5
本文 28.12 0.902 2 232.6

以上实验结果是在${\rho _a}$取0.2,${\rho _b}$取0.8下得到的,在图像重建中,不同的阈值取值会得到不同的重建效果,为进一步分析本文算法的性能,以Lena图为实验对象,仿真分析选用不同阈值的重建效果和重建时间,实验结果如图 2图 3所示。

图 2 ${\rho _a}$${\rho _b}$取值与SSIM关系
Fig. 2 Relation between two values and SSIM
图 3 ${\rho _a}$${\rho _b}$取值与时间关系
Fig. 3 Relation between two values and time

图 2图 3中Curve1表示了${\rho _a}$与SSIM和重建时间的关系(${\rho _b}$取0.8),可以看出${\rho _a}$取值增大,重建精度降低,但能较好地提高重建速度,由于取值过大将会除去对重建有一定贡献的原子,导致重建精度急剧下降,所以对于其取值不得超过0.4;Curve2表示了${\rho _b}$与SSIM和重建时间的关系(${\rho _a}$取0.2),可以看出${\rho _b}$取值增大,重建精度略有降低,但能提高一定的重建速度,但若取值过低,迭代过程中每次会抽取较多原子正则化处理加入支撑集,重建时间会增加过多,因此对其取值应高于0.6。而当两个阈值分别取到0和1时,本算法也就退化为经典的正交匹配追踪算法。

对于不同图像有不同的最佳阈值组合,经大量实验结果表明,在${\rho _a}$取0.2,${\rho _b}$取0.8下得到的结果具有普遍良好的重建效果,重建速度也有明显的提高。

4 结论

针对目前基于字典学习的超分辨率重建算法在重建效果和重建时间方面存在的问题,本文在字典学习的框架下作了改进,引入核方法,提高了字典原子选取的效率和准确度,进而降低了计算复杂度,同时也提高了重建质量。从实验结果可以看出,本文算法重建的图像相比于传统算法,重建图像更接近原图,边缘也更加平滑,锯齿现象更少;相对于字典学习的同类算法,本文的重建时间也更少。而如何选取合适的阈值组合,既保证重建质量,同时兼顾重建速度是本文需要继续讨论的问题。

参考文献

  • [1] Yu W J, Zhang M J. A mixed particle swarm optimization algorithm's application in image/video super-resolution reconstruction[C]//Proceedings of the 2nd International Conference on Image, Vision and Computing. Chengdu, China: IEEE, 2017: 526-530. [DOI: 10.1109/ICIVC.2017.7984611]
  • [2] Zhao S R, Lyu Z H, Liang H, et al. A mixed non-local prior model for image super-resolution reconstruction[J]. Chinese Journal of Electronics, 2017, 26(4): 778–783. [DOI:10.1049/cje.2016.06.024]
  • [3] Huang S Y, Sun J, Yang Y, et al. Multi-frame super-resolution reconstruction based on gradient vector flow hybrid field[J]. IEEE Access, 2017, 5: 21669–21683. [DOI:10.1109/ACCESS.2017.2757239]
  • [4] Wei D Y. Image super-resolution reconstruction using the high-order derivative interpolation associated with fractional filter functions[J]. IET Signal Processing, 2016, 10(9): 1052–1061. [DOI:10.1049/iet-spr.2015.0444]
  • [5] Chen H H, Jiang B L, Liu C, et al. Image super-resolution reconstruction based on residual error[J]. Journal of Image and Graphics, 2013, 18(1): 42–48. [陈华华, 姜宝林, 刘超, 等. 基于残差的图像超分辨率重建[J]. 中国图象图形学报, 2013, 18(1): 42–48. ] [DOI:10.11834/jig.20130105]
  • [6] Yang J C, Wang Z W, Lin Z, et al. Coupled dictionary training for image super-resolution[J]. IEEE Transactions on Image Processing, 2012, 21(8): 3467–3478. [DOI:10.1109/TIP.2012.2192127]
  • [7] Yang J C, Wright J, Huang T S, et al. Image super-resolution via sparse representation[J]. IEEE Transactions on Image Processing, 2010, 19(11): 2861–2873. [DOI:10.1109/TIP.2010.2050625]
  • [8] Zhu X, Liu L, Wang X X, et al. Super-resolution reconstruction via multiple sparse dictionary combined with sparse coding[C]//Proceedings of the IEEE International Conference on Information and Automation. Ningbo, China: IEEE, 2016: 1720-1725. [DOI: 10.1109/ICInfA.2016.7832095]
  • [9] Yang W M, Yuan T R, Wang W, et al. Single-image super-resolution by subdictionary coding and kernel regression[J]. IEEE Transactions on Systems, Man, and Cybernetics:Systems, 2017, 47(9): 2478–2488. [DOI:10.1109/TSMC.2016.2523947]
  • [10] Zeyde R, Elad M, Protter M. On single image scale-up using sparse-representations[C]//Proceedings of the 7th International Conference on Curves and Surfaces. Avignon, France: ACM, 2012, 6920: 711-730. [DOI: 10.1007/978-3-642-27413-8_47]
  • [11] Elad M. Sparse and Redundant Representations:From Theory to Applications in Signal and Image Processing[M]. Berlin, Germany: Springer-Verlag, 2010.
  • [12] Golts A, Elad M. Linearized kernel dictionary learning[J]. IEEE Journal of Selected Topics in Signal Processing, 2016, 10(4): 726–739. [DOI:10.1109/JSTSP.2016.2555241]
  • [13] Bryt O, Elad M. Compression of facial images using the K-SVD algorithm[J]. Journal of Visual Communication and Image Representation, 2008, 19(4): 270–282. [DOI:10.1016/j.jvcir.2008.03.001]
  • [14] Khademloo M, Rezghi M. Integrated single image super resolution based on sparse representation[C]//2015 International Symposium on Artificial Intelligence and Signal Processing. Mashhad, Iran: IEEE, 2015: 192-195. [DOI: 10.1109/AISP.2015.7123523]
  • [15] Yang J C, Wright J, Huang T, et al. Image super-resolution as sparse representation of raw image patches[C]//Proceedings of 2008 IEEE Conference on Computer Vision and Pattern Recognition. Anchorage, AK, USA: IEEE, 2008: 1-8. [DOI: 10.1109/CVPR.2008.4587647]
  • [16] Shi J, Wang X H. Image super-resolution reconstruction based on improved K-SVD dictionary-learning[J]. Acta Electronica Sinica, 2013, 41(5): 997–1000. [史郡, 王晓华. 基于改进K-SVD字典学习的超分辨率图像重构[J]. 电子学报, 2013, 41(5): 997–1000. ] [DOI:10.3969/j.issn.0372-2112.2013.05.026]
  • [17] Abolghasemi V, Ferdowsi S, Sanei S. Sparse multichannel source separation using incoherent K-SVD method[C]//Proceedings of 2011 IEEE Statistical Signal Processing Workshop (SSP). Nice, France: IEEE, 2011: 477-480. [DOI: 10.1109/SSP.2011.5967736]
  • [18] Wu D, Wang K M, Zhao Y X, et al. Stagewise regularized orthogonal matching pursuit algorithm[J]. Optics and Precision Engineering, 2014, 22(5): 1395–1402. [吴迪, 王奎民, 赵玉新, 等. 分段正则化正交匹配追踪算法[J]. 光学精密工程, 2014, 22(5): 1395–1402. ] [DOI:10.3788/OPE.20142205.1395]