一种改善图像边缘效果的POCS超分辨率重建方法

利用图像超分辨率重建(SRR)技术可以在现有成像系统基础上提高图像空间分辨力。凸集投影(POCS)是超分辨率重建的主流方法之一。对POCS算法进行了改进,具体改进体现在两个方面:(1)用可控核回归插值图像作为POCS重建的初始估计以提高初始估计图像的质量;(2)将POCS重建中使用的点扩散函数(PSF)由高斯核改为可控核以减少重建图像的边缘振荡效应。对所提出的算法进行了仿真,实验结果显示采用本文方法重建图像的边缘效果有了明显的改善。     

基于核回归正则项的图像超分辨率重建

讨论了图像成像的基本模型,并提出了一种基于调整核回归函数作为正则项的序列图像重建算法。该算法是对已经提出的核回归算法的改进,减少其在超分辨率图像重建时的运算量。而且在图像配准过程中针对图像间只存在平移和旋转变换,采用了基于矩形像素值的亚像素配准方法,以提高配准的速度和精度。利用此算法对序列图像进行重建仿真,并通过结论得出其在噪声严重的情况下具有更好的边缘保留特性。     

基于几何字典学习和耦合约束的超分辨率重建

传统的基于稀疏表示的超分辨率重建算法对所有图像块,应用单一冗余字典表示而不能反映不同几何结构类型图像块间的区别。针对这一问题,本文探索图像局部几何结构特性,提出一种基于结构特性聚类的几何字典学习和耦合约束的超分辨率重建方法。该方法首先对训练样本图像块进行几何特性聚类,然后应用K-SVD算法为每个聚类块联合训练得到高低分辨率字典。此外,在重建过程中引入局部可控核回归和非局部相似性耦合约束,以提高重建图像质量。实验结果表明,与单一字典超分辨率算法相比,本文方法重建图像边缘和细节部分明显改善,评价参数较大提高。     

基于2D-PCA特征描述的非负权重邻域嵌入人脸超分辨率重建算法

在基于邻域嵌入人脸图像的超分辨率重建算法中,训练和重建均在特征空间进行,因此,特征选择对算法性能具有较大影响。另外,算法模型对重建权重未加限定,导致负数权重出现而产生过拟合效应,使得重建人脸图像质量衰退。考虑到人脸图像的特征选择以及权重符号限定的重要作用,该文提出一种基于2维主成分分析(2D- PCA)特征描述的非负权重邻域嵌入人脸超分辨率重建算法。首先将人脸图像分成若干子块,利用K均值聚类获得图像子块的局部视觉基元,并利用得到的局部视觉基元对图像子块分类。然后,利用2D-PCA对每一类人脸图像子块提取特征,并建立高、低分辨率样本库。最后,在重建过程中使用新的非负权重求解方法求取权重。仿真实验结果表明,相比其他基于邻域嵌入人脸超分辨率重建方法,所提算法可有效提高权重的稳定性,减少过拟合效应,其重建人脸图像具有较好的主客观质量。     

基于多核支持向量回归的光谱反射率重建方法

针对传统方法重建光谱反射率过程中未考虑多光谱训练数据维度高、冗余大的特点,导致重建精度低、重建模型学习能力和泛化能力较差的问题,提出了多核支持向量回归的光谱反射率重建方法。首先综合全局、局部核函数的特点引入柯西核函数与多项式核函数的乘积作为多核核函数,然后使用试凑法从训练样本中获取提高模型性能的参数。最后使用多核支持向量回归模型对测试样本进行反射率重建,通过光谱误差及适应度等进行评价。实验结果表明:与伪逆、单核支持向量回归法相比,本文重建方法的光谱误差值降低了0.4～0.785,决策系数提高了2.84～5.27%,平均适应度系数值提高了2%～3%。本文方法在颜色复制中重建精度高、色差较小,满足人眼视觉可容忍范围内。     

基于改进邻域嵌入和结构自相似性的超分辨率重建方法

现有的基于压缩感知的超分辨率重建模型需要对 高分辨率图像进行初始估计,而初始估计的准确与 否直接影响图像重建的质量与迭代次数。针对此问题,本文引入非局部均值正则项以改进邻 域嵌入方法, 从而获得更加准确的高频初始估计;同时利用低分辨率图像的局部自相似性和多尺度结构相 似性构建约 束项,从而提出了一种基于改进的邻域嵌入和结构自相似性的超分辨率重建方法,充分结合 两者的优势, 增强了先验估计的表达能力。实验结果表明,相较于现有算法,本文提出的算法在客观评价 指标和主观视觉质量上均有显著提高。     

图像融合在空间目标三维重建中的应用

图像融合可以获取目标更加丰富的层次和细节信息，有利于对探测目标信息的有效地获取，在包括空间目标的三维重建等应用中有着重要意义。针对空间目标的宽动态范围提出了一种多次曝光的图像融合方法，利用信息熵的非线性压缩判定图像融合权重，并引入双边滤波残差加强弱纹理的分配权重，有效地增加了图像的特征信息，提高了三维重建点云的数量。利用提出的融合方法开展了空间目标模拟成像试验，采用融合的图像对目标三维重建，并与多种不同曝光程度以及采用其他融合图像的方法进行了对比，提出的方法得到的重建点云数量相对恰当曝光状态提高了35%，重建结果优于其他方法。结果表明:将图像融合引入到三维重建中，能有效地加强了重建图像信息，避免了光照条件对目标三维重建的不利影响，获得较高质量的重建效果，该方法可以很好地应用到基于图像序列的空间目标三维重建应用中。     

基于加权Boosting的核偏最小二乘图像超分辨率重建

核偏最小二乘(KPLS)算法对每个图像块选用全部主元成分进行图像重建,导致图像超分辨率算法的计算量大。兼顾图像重建质量和时间效率,该文提出一种加权Boosting的图像超分辨率重建算法。为自适应地选取每个图像块主元成分的最佳数目,利用加权Boosting原理对KPLS回归预测量进行补偿,推导给出补偿权重系数的数学表达式。讨论不同Boosting阈值情况下的重建性能,在合适的下,选取出主元成分的最佳数目m更好地满足KPLS回归模型的精度要求。实验结果表明,该文算法的超分辨率重建质量优于传统算法。     

基于边缘方向扩散系数正则化的文本图像超分辨率算法

针对文档图像超分辨率重建问题,根据传统双边全变差(BTV)超分辨率算法,提出一种自适应约束的BTV正则化文档图像超分辨率算法。该算法通过引入一个图像的局部邻域残差均值,以区分当前像素点属于平滑区域还是边缘区域,然后利用垂直边缘方向和边缘方向扩散性的不同,产生自适应权重矩阵。最后通过代价函数求出迭代公式,最终实现文本图像的超分辨率重建。与相关的文档图像超分辨率方法相比较,提出的方法在视觉图像质量和字符识别精度方面均得到了显著的改善。     

基于MDCGAN的裂缝样本扩充及识别研究

针对裂缝图像获取困难导致的样本少、传统数据扩充方法提升样本特征空间能力不足等问题,提出了一种基于改进深度卷积生成对抗网络(MDCGAN)的裂缝样本扩充方法。首先对数据集进行预处理,利用滑窗法进行数据降维和清洗;其次优化激活函数,提高生成特征的多样性,同时引入谱归一化进行权重标准化提升网络结构的稳定性,以生成高质量的裂缝数据集;最后,利用改进的Alexnet网络对扩充后的混合样本集进行特征提取并分类识别。结果表明,MDCGAN网络数据增强性能与传统扩充方法相比均有明显提高,适用于扩充裂缝图像。     

On Kernel Selection of Multivariate Local Polynomial Modelling and its Application to Image Smoothing and Reconstruction

This paper studies the problem of adaptive kernel selection for multivariate local polynomial regression (LPR) and its application to smoothing and reconstruction of noisy images. In multivariate LPR, the multidimensional signals are modeled locally by a polynomial using least-squares (LS) criterion with a kernel controlled by a certain bandwidth matrix. Based on the traditional intersection confidence intervals (ICI) method, a new refined ICI (RICI) adaptive scale selector for symmetric kernel is developed to achieve a better bias-variance tradeoff. The method is further extended to steering kernel with local orientation to adapt better to local characteristics of multidimensional signals. The resulting multivariate LPR method called the steering-kernel-based LPR with refined ICI method (SK-LPR-RICI) is applied to the smoothing and reconstruction problems in noisy images. Simulation results show that the proposed SK-LPR-RICI method has a better PSNR and visual performance than conventional LPR-based methods in image processing.     

基于迭代交替优化的图像盲超分辨率重建

基于深度卷积神经网络的图像超分辨率重建算法通常假设低分辨率图像的降质是固定且已知的,如双3次下采样等,因此难以处理降质(如模糊核及噪声水平)未知的图像。针对此问题,该文提出联合估计模糊核、噪声水平和高分辨率图像,设计了一种基于迭代交替优化的图像盲超分辨率重建网络。在所提网络中,图像重建器以估计的模糊核和噪声水平作为先验信息,由低分辨率图像重建出高分辨率图像;同时,综合低分辨率图像和估计的高分辨率图像,模糊核及噪声水平估计器分别实现模糊核和噪声水平的估计。进一步地,该文提出对模糊核/噪声水平估计器及图像重建器进行迭代交替的端对端优化,以提高它们的兼容性并使其相互促进。实验结果表明,与IKC, DASR, MANet, DAN等现有算法相比,提出方法在常用公开测试集(Set5, Set14, B100, Urban100)及真实场景图像上都取得了更优的性能,能够更好地对降质未知的图像进行重建;同时,提出方法在参数量或处理效率上也有一定的优势。     

A High Quality Depth Map Upsampling Method Robust to Misalignment of Depth and Color Boundaries

In recent years, fusion camera systems that consist of color cameras and Time-of-Flight (TOF) depth sensors have been popularly used due to its depth sensing capability at real-time frame rates. However, captured depth maps are limited in low resolution compared to the corresponding color images due to physical limitation of the TOF depth sensor. Most approaches to enhancing the resolution of captured depth maps depend on the implicit assumption that when neighboring pixels in the color image have similar values, they are also similar in depth. Although many algorithms have been proposed, they still yield erroneous results, especially when region boundaries in the depth map and the color image are not aligned. We therefore propose a novel kernel regression framework to generate the high quality depth map. Our proposed filter is based on the vector pointing similar pixels that represents the unit vector toward similar neighbors in the local region. The vectors are used to detect misaligned regions between color edges and depth edges. Unlike conventional kernel regression methods, our method properly handles misaligned regions by introducing the numerical analysis of the local structure into the kernel regression framework. Experimental comparisons with other data fusion techniques prove the superiority of the proposed algorithm.     

Computed tomographic imaging of the coronary arterial tree-use of local tomography

It has been demonstrated previously (see E.L. Ritman and A.A. Bove, in State of the Art in Quantitative Coronary Arteriography, p.67-78, 1986) that coronary artery anatomy can be visualized using high-speed, volume-scanning X-ray CT (computed tomography). In the current study it is demonstrated that local image reconstruction (i.e. the reconstruction kernel is ~2(+) mm long), as distinct from more conventional global image reconstruction (i.e. 200(+) mm kernel length), has the advantage of reducing the need for operator interactive image processing. In addition, the local reconstruction algorithm eliminates the need for recording the X-ray projection data over the full transaxial extent of the thorax because it requires only the X-ray attenuation data over the region of the heart. This latter aspect reduces the dynamic range requirements for the sensors and could reduce total X-ray exposure.     

基于分块LBP和鲁棒核编码的人脸识别

针对人脸识别中的遮挡和姿态偏转等问题,提出了一种基于分块LBP和鲁棒核编码(Robust Kernel Coding,RKC)的人脸识别算法,简称LBP-RKC算法.该算法首先对人脸图像进行多级分块的LBP特征提取,得到图像的每一块统计直方图特征.然后,将特征投影到核空间中,在核空间中建立一个鲁棒的回归模型来处理图像中的异常值,并利用迭代重加权算法求解该模型.最后,计算测试样本的每一块核表示重构残差并进行分类识别.实验表明,提出的LBP-RKC算法在处理遮挡、姿态偏转等人脸问题时能取得很好的识别效果,同时算法效率较高.     

Least-Square NUFFT Methods Applied to 2-D and 3-D Radially Encoded MR Image Reconstruction

Radially encoded MRI has gained increasing attention due to its motion insensitivity and reduced artifacts. However, because its samples are collected nonuniformly in the $k$-space, multidimensional (especially 3-D) radially sampled MRI image reconstruction is challenging. The objective of this paper is to develop a reconstruction technique in high dimensions with on-the-fly kernel calculation. It implements general multidimensional nonuniform fast Fourier transform (NUFFT) algorithms and incorporates them into a $k$-space image reconstruction framework. The method is then applied to reconstruct from the radially encoded $k$-space data, although the method is applicable to any non-Cartesian patterns. Performance comparisons are made against the conventional Kaiser–Bessel (KB) gridding method for 2-D and 3-D radially encoded computer-simulated phantoms and physically scanned phantoms. The results show that the NUFFT reconstruction method has better accuracy–efficiency tradeoff than the KB gridding method when the kernel weights are calculated on the fly. It is found that for a particular conventional kernel function, using its corresponding deapodization function as a scaling factor in the NUFFT framework has the potential to improve accuracy. In particular, when a cosine scaling factor is used, the NUFFT method is faster than KB gridding method since a closed-form solution is available and is less computationally expensive than the KB kernel (KB griding requires computation of Bessel functions). The NUFFT method has been successfully applied to 2-D and 3-D in vivo studies on small animals.      

Total least square kernel regression

In this paper, we study the problem of robust image fusion in the context of multi-frame super-resolution. Given multiple aligned noisy low-resolution images, image fusion produces a new image on a high-resolution grid. Recently, kernel regression is presented as a powerful image fusion technique. However, in the presence of registration errors, the performance of kernel regression is quite poor. Therefore, we present a new kernel regression method that takes these registration errors into account. Instead of the ordinary least square metric, we employ the total least square metric, which allows for spatial perturbations of the image samples. We show in our experiments that our method is more robust to noise and/or registration errors compared to the traditional kernel regression algorithm.