Image super resolution based on residual dense CNN and guided filters

Convolutional neural networks (CNNs) have recently made impressive results for image super-resolution (SR). Our goal is to introduce a new image SR framework rely on a CNN. In this paper, the input image is decomposed into luminance channel and chromatic channels. A designed network based on a residual dense network is introduced to extract the hierarchical features from luminance part. The bicubic interpolation is simply used to upscale low resolution (LR) chromatic channels. However, this step degrades the chromatic channels. To tackle this issue, the SR reconstructed luminance channel is applied as the reference image in guided filters to promote the interpolated chromatic channels. Guided filters technique has ability to retain sharp edges and fine details from the reference image and carry them to the target images. Extensive experiments on several commonly used image SR testing datasets demonstrate that our framework has the ability to extract features and outperforms existing well-known techniques for image SR by LR image into the high resolution (HR) image efficiently.

     

Effects of the deep learning-based super-resolution method on thermal image classification applications

Thermal imaging can be used in many sectors such as public security, health, and defense in image processing. However, thermal imaging systems are very costly, limiting their use, especially in the medical field. Also, thermal camera systems obtain blurry images with low levels of detail. Therefore, the need to improve their resolution has arisen. Here, super-resolution techniques can be a solution. Developments in deep learning in recent years have increased the success of super-resolution (SR) applications. This study proposes a new deep learning-based approach TSRGAN model for SR applications performed on a new dataset consisting of thermal images of premature babies. This dataset was created by downscaling the thermal images (ground truth) of premature babies as traditional SR studies. Thus, a dataset consisting of high-resolution (HR) and low-resolution (LR) thermal images were obtained. SR images created due to the applications were compared with LR, bicubic interpolation images, and obtained SR images using state-of-the-art models. The success of the results was evaluated using image quality metrics of peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM). The results show that the proposed model achieved the second-best PSNR value and the best SSIM value. Additionally, a CNN-based classifier model was developed to perform task-based evaluation, and classification applications were carried out separately on LR, HR, and reconstructed SR image sets. Here, the success of classifying unhealthy and healthy babies was compared. This study showed that the classification accuracy of SR images increased by approximately 5% compared to the classification accuracy of LR images. In addition, the classification accuracy of SR thermal images approached the classification accuracy of HR thermal images by about 2%. Therefore, with the approach proposed in this study, it has been proven that LR thermal images can be used in classification applications by increasing their resolution. Thus, widespread use of thermal imaging systems with lower costs in the medical field will be achieved.

     

Image super-resolution via two coupled dictionaries and sparse representation

In image processing, the super-resolution (SR) technique has played an important role to perform high-resolution (HR) images from the acquired low-resolution (LR) images. In this paper, a novel technique is proposed that can generate a SR image from a single LR input image. Designed framework can be used in images of different kinds. To reconstruct a HR image, it is necessary to perform an intermediate step, which consists of an initial interpolation; next, the features are extracted from this initial image via convolution operation. Then, the principal component analysis (PCA) is used to reduce information redundancy after features extraction step. Non-overlapping blocks are extracted, and for each block, the sparse representation is performed, which it is later used to recover the HR image. Using the quality objective criteria and subjective visual perception, the proposed technique has been evaluated demonstrating their competitive performance in comparison with state-of-the-art methods.     

Learning graph-constrained cascade regressors for single image super-resolution

Learning cascade regression has been shown an effective strategy to further enhance the perceptual quality of resulted high-resolution (HR) images. However, previous cascade regression-based SR methods have two obvious weaknesses: (1)edge structures cannot be preserved well when applying texture features to represent low-resolution (LR) images, and (2)the local manifold structures spanned by the LR-HR feature spaces cannot be revealed by the learned local linear mappings. To alleviate the aforementioned problems, a novel example regression-based super-resolution (SR) approach called learning graph-constrained cascade regressors (LGCCR) is presented, which learns a group of multi-round residual regressors in a unique way. Specifically, we improve the edge preservation capability by synthesizing the whole HR image rather than local image patches, which facilitates to extract the edge features to represent LR images. Moreover, we utilize a graph-constrained regression model to build the local linear regressors, where each local linear regressor responds to an anchored atom in the learned over-complete dictionary. Both quantitative and qualitative quality evaluations on seven benchmark databases indicate the superiority of the proposed LGCCR-based SR approach in comparing with other state-of-the-art SR predecessors.

     

结合支持向量回归和图像自相似的单幅图像超分辨率算法

目的 基于学习的单幅图像超分辨率算法是借助实例训练库由一幅低分辨率图像产生高分辨率图像。提出一种基于图像块自相似性和对非线性映射拟合较好的支持向量回归模型的单幅超分辨率方法,该方法不需使用外部图像训练库。方法 首先根据输入的低分辨率图像建立图像金字塔及包含低/高分辨率图像块对的集合;然后在低/高分辨率图像块对的集合中寻找与输入低分辨率图像块的相似块,利用支持向量回归模型学习这些低分辨率相似块和其对应的高分辨率图像块的中心像素之间的映射关系,进而得到未知高分辨率图像块的中心像素。结果 为了验证本文设计算法的有效性,选取结构和纹理不同的7幅彩色高分辨率图像,对其进行高斯模糊的2倍下采样后所得的低分辨率图像进行超分辨率重构,与双三次插值、基于稀疏表示及基于支持向量回归这3个超分辨率方法重建的高分辨率图像进行比较,峰值信噪比平均依次提升了2.37 dB、0.70 dB和0.57 dB。结论 实验结果表明,本文设计的算法能够很好地实现图像的超分辨率重构,特别是对纹理结构相似度高的图像具有更好的重构效果。     

递归式多阶特征融合图像超分辨率算法

目的 近年来,卷积神经网络在解决图像超分辨率的问题上取得了巨大成功,不同结构的网络模型相继被提出。通过学习,这些网络模型对输入图像的特征进行抽象、组合,进而建立了从低分辨率的输入图像到高分辨率的目标图像的有效非线性映射。在该过程中,无论是图像的低阶像素级特征,还是高阶各层抽象特征,都对像素间相关性的挖掘起了重要作用,影响着目标高分辨图像的性能。而目前典型的超分辨率网络模型,如SRCNN（super-resolution convolutional neural network）、VDSR（very deep convolutional networks for super-resolution）、LapSRN（Laplacian pyramid super-resolution networks）等,都未充分利用这些多层次的特征。方法 提出一种充分融合网络多阶特征的图像超分辨率算法：该模型基于递归神经网络,由相同的单元串联构成,单元间参数共享;在每个单元内部,从低阶到高阶的逐级特征被级联、融合,以获得更丰富的信息来强化网络的学习能力;在训练中,采用基于残差的策略,单元内使用局部残差学习,整体网络使用全局残差学习,以加快训练速度。结果 所提出的网络模型在通用4个测试集上,针对分辨率放大2倍、3倍、4倍的情况,与深层超分辨率网络VDSR相比,平均分别能够获得0.24 dB、0.23 dB、0.19 dB的增益。结论 实验结果表明,所提出的递归式多阶特征融合图像超分辨率算法,有效提升了性能,特别是在细节非常丰富的Urban100数据集上,该算法对细节的处理效果尤为明显,图像的客观质量与主观质量都得到显著改善。     

多通道卷积的图像超分辨率方法

目的 超分辨率技术在实际生活中具有较为广泛的应用。经典的基于卷积神经网络的超分辨率（SRCNN）方法存在重建图像纹理结构模糊以及网络模型训练收敛过慢等问题。针对这两个问题,在SRCNN的基础上,提出一种多通道卷积的图像超分辨率（MCSR）算法。方法 通过增加残差链接,选择MSRA初始化方法对网络权值进行初始化,加快模型收敛;引入多通道映射提取更加丰富的特征,使用多层3×3等小卷积核代替单层9×9等大卷积核,更加有效地利用特征,增强模型的超分辨率重构效果。结果 MCSR迭代4×10⁶次即可收敛,在Set5与Set14数据集上边长放大3倍后的平均峰值信噪比分别是32.84 dB和29.28 dB,与SRCNN相比提升显著。结论 MCSR收敛速度更快,并且可以生成轮廓清晰的高分辨率图像,超分辨率效果更加优秀。     

基于低分辨率图像融合的超分辨率恢复方法

提出一种基于多帧低分辨图像融合的超分辨率图像恢复算法。将多帧低分辨图像融合成一帧与高分辨图像分辨率一致的图像,并对其中一幅低分辨图像插值,形成迭代恢复算法的初始值,在此基础上以Tikhonov正则化方法求解原始高分辨图像。分析和实验结果表明,该算法具有良好的鲁棒性,并且计算速度较快。     

Locally affine patch mapping and global refinement for image super-resolution

This paper deals with the super-resolution (SR) problem based on a single low-resolution (LR) image. Inspired by the local tangent space alignment algorithm in [16] for nonlinear dimensionality reduction of manifolds, we propose a novel patch-learning method using locally affine patch mapping (LAPM) to solve the SR problem. This approach maps the patch manifold of low-resolution image to the patch manifold of the corresponding high-resolution (HR) image. This patch mapping is learned by a training set of pairs of LR/HR images, utilizing the affine equivalence between the local low-dimensional coordinates of the two manifolds. The latent HR image of the input (an LR image) is estimated by the HR patches which are generated by the proposed patch mapping on the LR patches of the input. We also give a simple analysis of the reconstruction errors of the algorithm LAPM. Furthermore we propose a global refinement technique to improve the estimated HR image. Numerical results are given to show the efficiency of our proposed methods by comparing these methods with other existing algorithms.     

Efficient single image super-resolution via graph-constrained least squares regression

We explore in this paper an efficient algorithmic solution to single image super-resolution (SR). We propose the gCLSR, namely graph-Constrained Least Squares Regression, to super-resolve a high-resolution (HR) image from a single low-resolution (LR) observation. The basic idea of gCLSR is to learn a projection matrix mapping the LR image patch to the HR image patch space while preserving the intrinsic geometric structure of the original HR image patch manifold. Even if gCLSR resembles other manifold learning-based SR methods in preserving the local geometric structure of HR and LR image patch manifolds, the innovation of gCLSR lies in that it preserves the intrinsic geometric structure of the original HR image patch manifold rather than the LR image patch manifold, which may be contaminated by image degeneration (e.g., blurring, down-sampling and noise). Upon acquiring the projection matrix, the target HR image can be simply super-resolved from a single LR image without the need of HR-LR training pairs, which favors resource-limited applications. Experiments on images from the public database show that gCLSR method can achieve competitive quality as state-of-the-art methods, while gCLSR is much more efficient in computation than some state-of-the-art methods.     

压缩图像空时自适应正则化超分辨率重建

所谓超分辨率(SR)技术就是由低分辨率(LR)图像序列来重建高分辨率(HR)图像的技术,而基于压缩图像的SR技术正成为当前研究的热点。为了提高压缩图像的重建质量,在正则化理论的基础上,通过利用比特流中的信息,提出了一种新颖的空时自适应超分辨率重建算法,该算法先利用正则化代价函数控制时域数据和空域先验信息之间的平衡,使正则化参数在SR重建过程中得到自适应地调整,然后利用迭代梯度下降法进行超分辨率重建。仿真实验表明,该自适应算法比采用传统算法重建的图像的主、客观质量有一定的提高,适合压缩图像的应用。     

基于特征图注意力机制的图像超分辨率重建

图像超分辨率重建中的高频分量通常包含较多轮廓、纹理等细节信息,为更好地处理特征图中的高频分量与低频分量,实现自适应调整信道特征,提出一种基于特征图注意力机制的图像超分辨重建网络模型。利用特征提取块提取原始低分辨率图像中的特征信息,基于多个结合特征图注意力机制的信息提取块,通过特征信道之间的相互依赖性自适应调整信道特征,以恢复更多细节信息。在此基础上利用重建模块重建出不同尺度的高分辨率图像。在Set5数据集上的实验结果表明,与基于双三次插值的重建模型相比,该模型能够有效提升图像的视觉效果,且峰值信噪比与结构相似度分别提高了3.92 dB和0.056。     

统一最小二乘规则的单幅图像超分辨算法

基于机器学习的超分辨方法是一个很有发展前景的单幅图像超分辨方法,稀疏表达和字典学习是其中的研究热点。针对比较耗时的字典训练与恢复精度不高图像重建,从减小低分辨率(LR)和高分辨率(HR)特征空间之间差异性的角度提出了一种使用迭代最小二乘字典学习算法(ILS-DLA),并使用锚定邻域回归(ANR)进行图像重建的单幅图像超分辨算法。迭代最小二乘法的整体优化过程极大地缩短了低分辨字典/高分辨字典的训练时间,它采用了与锚定邻域回归相同的优化规则,有效地保证了字典学习和图像重建在理论上的一致性。实验结果表明,所提算法的字典学习效果比K-均值奇异值分解(K-SVD)和Beta过程联合字典学习(BPJDL)等算法更高效,图像重建的效果也优于许多优秀的超分辨算法。     

Video super-resolution network using detail component extraction and optical flow enhancement algorithm

The video super-resolution (SR) task refers to the use of corresponding low-resolution (LR) frames and multiple neighboring frames to generate high-resolution (HR) frames. Existing deep learning-based approaches usually utilize LR optical flow for video SR tasks. However, the accuracy of LR optical flow is not enough to recover the fine detail part. In this paper, we propose a video SR network that uses optical flow SR and optical flow enhancement algorithms to provide accurate temporal dependency. And extract the detail component of LR adjacent frames as supplementary information for accurate feature extraction. Firstly, the network infers HR optical flow from LR optical flow, and uses the optical flow enhancement algorithm to enhance HR optical flow. Then the processed HR optical flows are used as the input of the motion compensation network. Secondly, we extract detail component to reduce the error caused by motion compensation based on optical flow. Finally, the SR results are generated through the SR network. We perform comprehensive comparative experiments on two datasets: Vid4 and DAVIS. The results show that, compared with other state-of-the-art methods, the proposed video SR method achieves the better performance.

     

K均值聚类和支持向量数据描述的图像超分辨率算法

目的 为了提高图像超分辨率算法对数据奇异点的鲁棒性,提出一种采用K均值聚类和支持向量数据描述的图像超分辨率重建算法(Kmeans-SVDD)。方法 训练过程:首先用K均值聚类算法将训练图像的近似子带划分为若干类,然后用支持向量数据描述去除每类数据的奇异点,最后在小波域内用主成分分析训练近似子带和细节子带字典。测试过程:根据同一场景高低分辨率图像近似子带相似这一现象,首先将待重建低分辨率测试图像的近似子带作为相应高分辨率测试图像的近似子带,然后由训练得到的字典恢复出高分辨率测试图像的细节子带,最后通过逆小波变换得到高分辨率测试图像。结果 相比于当前双三次插值、Zeyde、ANR与Kmeans-PCA算法,Kmeans-SVDD算法重建的高分辨率测试图像的平均峰值信噪比依次提高了1.82 dB、0.37 dB、0.30 dB、0.15 dB。结论 通过大量实验发现,在字典训练之前加入SVDD过程可以去除离群点,提高字典质量。在小波域中将各频带分开重建,可避免低频图像中包含的不可靠高频信息对超分辨率结果的影响,从而恢复出可靠的高频信息。     

基于残差的图像超分辨率重建

提出一种基于图像残差的超分辨率重建算法.以原高分辨率图像与插值放大后图像之间的图像残差与低分辨率图像样本特征作为样本对,对其进行K均值分类,并对每类样本对采用KSVD(K-singular value decomposition)方法进行训练获得高、低分辨率字典对,然后根据测试样本与类中心的欧氏距离选择字典对,以与测试样本相近的多个类别所重建的结果加权获得图像残差,并结合低分辨率图像的插值结果获得高分辨率图像.实验结果表明,提出的方法具有更高的重建质量,且采用训练样本分类和相近类别的重建结果的加权和有利于提高图像重建质量.     

压缩视频的正则化投影超分辨率重建

压缩视频超分辨率（SR）技术利用压缩后的低分辨率（LR）图像序列来重建高分辨率（HR）图像的技术,是当前视频超分辨率技术研究的热点。在正则化理论和凸集投影理论的基础上,利用比特流中的量化信息,提出了一种正则化投影超分辨率重建算法;通过正则化代价函数引入图像序列的时间域和空间域的先验信息,使用迭代梯度下降算法对正则化代价函数求解得到重建图像,最后利用凸集投影算法对求得的估计图像进行DCT域投影重建。仿真实验结果表明,该自适应算法较传统算法,其重建图像的主、客观质量有一定的提高,适合压缩图像的应用。     

New edge-adaptive image interpolation using anisotropic Gaussian filters

Edge-adaptive interpolation is an effective technique to obtain a sharp high-resolution (HR) image from a low-resolution (LR) image. In this paper, we propose a new image interpolation method using edge-adaptive anisotropic Gaussian filters (AGFs). In the proposed method, the edge information of the LR image, which is the edge orientation and its energy, is first estimated using the modified Leung–Malik (LM) filter bank. Then, the edge information of the LR image is converted into that of the HR image by using a mapping function. Based on the HR edge information, edge-adaptive interpolation is performed using the AGF with variable shape and direction. Simulation results show that the proposed method objectively and subjectively outperforms conventional edge-based methods.     

结合深度学习的单幅遥感图像超分辨率重建

目的 克服传统遥感图像超分辨率重建方法依赖同一场景多时相图像序列且需预先配准等缺点,解决学习法中训练效率低和过拟合问题,同时削弱插值操作后的块效应,增强单幅遥感图像超分辨率重建效果。方法 首先构造基于四层卷积的深度神经网络结构,并在结构中前三层卷积后添加参数修正线性单元层和局部响应归一化层进行优化,经过训练得到遥感图像超分辨率重建模型,其次,对多波段遥感图像的亮度空间进行双三次插值,然后使用该模型对插值结果进行重建,并在亮度空间重建结果指导下,使用联合双边滤波来提升其色度空间边缘细节。结果 应用该方法对实验遥感图像进行2倍、3倍、4倍重建时在无参考指标上均优于对比方法,平均清晰度提升约2.5个单位,同时取得了较好的全参考评价结果,在2倍重建时峰值信噪比较传统插值法提升了约2 dB,且平均训练效率较其他学习法提升3倍以上,所得遥感图像重建结果在目视效果上更加细致、自然。结论 实验结果表明,本文设计的网络抗过拟合能力强、训练效率高,重建时针对单幅遥感图像,无需依赖图像序列且不受波段影响,重建结果细节表现较好,具有较强的普适性。     

基于非线性最小二乘的图像自适应SR重建以及运动估计

图像超分辨率（SR）重建是利用数字信号处理技术由一系列低分辨率观测图像得到高分辨率图像。为了扩展SR技术的应用范围,提出了一种同时进行图像超分辨率重建和全局运动估计的方法。该方法首先基于最大后验概率(MAP)给出了图像SR重建和运动估计框架,该框架不仅考虑了前后两次迭代所得的HR图像差值对最终重建图像的影响,而且引入了不同LR图像对重建图像的重要性权值,使得算法具有自适应性;然后将总体框架转换为图像SR重建模型和运动估计模型;最后基于非线性最小二乘法对模型进行优化求解,得出了SR重建图像及其全局运动域。实验表明,该方法不仅图像重建效果良好,并有着良好的收敛性。