Super-resolution imaging: use of zoom as a cue

In this paper, we propose a novel technique for super-resolution imaging of a scene from observations at different zoom levels. Given a sequence of images with different zoom factors of a static scene, the problem is to obtain a picture of the entire scene at a resolution corresponding to the most zoomed image in the scene. We not only obtain the super-resolved image for known integer zoom factors, but also for unknown arbitrary zoom factors. We model the super-resolution image as a Markov random field (MRF) and a maximum a posteriori (MAP) estimation method is used to derive a cost function which is then optimized to recover the high-resolution field. The entire observation conforms to the same MRF, but is viewed at the different resolution pyramid. Since there is no relative motion between the scene and the camera, as is the case with most of the super-resolution techniques, we do away with the correspondence problem. Results of the experimentation on real data are presented.     

文本图像的双边总变分超分辨率恢复方法

超分辨率图像复原的目的是从几幅包含有相同观测区域的图像序列中重建更高分辨率的图像。一般的超分辨率算法没有考虑文本图像的分段连续性，因此有可能丢失笔画边缘处的细节信息。双边总变分正则算法具有很好的边缘保持能力，因此可以应用到文本图像的超分辨率恢复。本文根据汉字的字符结构特点，对双边总变分正则算法进行了改进，根据汉字的笔画组成特征。提出新的正则项。该方法由于充分地考虑了汉字文本图像的结构特征，因此能够很好地保持汉字的笔画的尖锐边缘。实验也证明了使用本文所提出的BTVCH算法获得的超分辨率文本图像可以获得最高的字符识别率。     

基于ARNOLDI过程的参数估计

本文提出在超分辨率复原中使用基于Amoldi过程来高效计算正则化参数的方法。通过Arnoldi过程分解，该方法将大型稀疏系统矩阵投影到Krylov子空间上并表达成一个小型稠密的Hessenberg矩阵。给出了利用Hessenberg矩阵简化超分辨率复原中解计算的公式。推导了快速计算L曲线的定理。该方法可减少正则化参数的计算代价。     

基于非局部POCS的超分辨率图像重建

图像获取过程中,受成像系统的影响,无法获取原始场景中所有的信息。超分辨率图像重建技术就是在不改变成像系统的前提下,提高图像质量。POCS(凸集投影算法)可以利用多帧低分辨率图像重建一帧高分辨率图像。然而传统的POCS算法通常会产生"锯齿"边缘。在自然图像中,会存在许多的相似边缘结构。利用局部相似性的结构,可以有效地消除"锯齿"边缘。因此提出一种基于非局部POCS的超分辨率图像重建算法,以有效锐化图像边缘,提高图像的视觉感观。     

Investigation of multi-frame image super-resolution based on adaptive self-learning methods

The goal of this paper is to introduce and demonstrate a new high-performance super-resolution (SR) method for multi-frame images. By combining learning-based and reconstruction-based SR methods, this paper proposes a multi-frame image super-resolution method based on adaptive self-learning. Using the adaptive self-learning method and recovery of high-frequency edge information, an initial high-resolution (HR) image containing effective texture information is obtained. The edge smoothness prior is then used to satisfy the global reconstruction constraint and enhance the quality of the HR image. Our results indicate that this method achieves better performance than several other methods for both simulated data and real-scene images.     

基于压缩感知和SVR的自学习单幅图像超分辨率重建

针对依靠外部图像库的超分辨率(SR)重建算法训练耗时长、容易出现错误高频细节的问题,提出了一种基于压缩感知(CS)理论和支持向量回归(SVR)的单幅图像超分辨率重建方法。对降质图像本身训练SVR模型,充分挖掘图像自身的自相似特点。训练过程中先对输入图像边缘进行检测并对图像块进行分类,然后稀疏编码图像块,再根据图像的标签向量和稀疏表示矩阵训练得到SVR模型,并在测试过程中利用该模型预测高分辨率(HR)图像。实验结果表明,与基于外部库方法重建图像的方法相比,该算法所得结果的细节更加真实;与双三次插值方法相比该算法所得结果的边缘更加清晰。     

基于局部超分辨重建的高精度SAR图像水域分割方法

合成孔径雷达(SAR)图像水域分割在水资源调查、灾害监测等领域具有重要意义。针对中低分辨率星载SAR图像水域提取精度不足的难题,该文融合基于轻量级残差卷积神经网络(CNN)的图像超分辨率重建技术和传统SAR图像水域分割技术的优点,提出了一种基于局部超分辨重建的SAR图像水域分割方法,显著提升了SAR图像水域分割的精度。为了验证上述方法的有效性,该文以南水北调中线工程水源地丹江口水库为应用对象,基于国产高分三号(GF-3)卫星的8 m分辨率标准条带(SS)模式图像和欧空局Sentinel-1卫星20 m分辨率干涉宽幅(IW)模式图像,开展了水域分割的实验验证和精度评估工作。实验结果表明,该文所提方法可在中低分辨率SAR图像中获取更精确的水域分割结果,其水域分割性能较传统方法有大幅提升。      

基于SRCNN和SSD网络的小目标检测方法

针对实时目标检测SSD(Sing shot multibox detector)算法对小目标检测能力弱的问题,提出一种提高特征图分辨率的超分辨率重建SRCNN(Super Resolution Convolutional Neural Networks)设计策略。改进算法是在SSD基础网络VGG_16网络的conv4_3卷积层上进行的,把conv4_3卷积层产生的特征图通过SRCNN网络进行超分辨重建以提高conv4_3卷积层的特征图分辨率。然后再利用超分辨重建后的特征图和原特征图一起为小目标检测提供所需要的特征。实验表明上述设计方法相比于原经典SSD算法具有更高的检测精度和检测能力,以及在小目标检测上的效果更加明显。     

LD夜视成像探测中基于PSF的超分辨率重建技术研究

夜视目标识别在视频监控中起着非常重要的作用。从早期的红外热成像发展至今,激光夜视成像系统被广泛的应用。它不仅受到大气吸收和散射特性的影响,还与光学系统中的透镜及传感器的衍射极限有关。在既有成本下,通过图像复原和超分辨率重建,可以尽可能的提高夜视成像的成像质量。很多学者的研究结果表明,成像系统的点扩散函数作为先验知识,可以提高图像复原的质量。因此,本文在最大后验概率框架下,实现点扩散函数与凸集投影超分辨率重建相结合,并应用到半导体激光夜视成像探测系统中,客观的图像质量评价标准被用来决定迭代次数。目标识别的实验结果表明,用本文提出的重建方法可以有效提高半导体激光夜视成像的范围和成像质量。     

Super-resolution image reconstruction techniques: Trade-offs between the data-fidelity and regularization terms

Stochastic regularized methods are quite advantageous in super-resolution (SR) image reconstruction problems. In the particular techniques, the SR problem is formulated by means of two terms, the data-fidelity term and the regularization term. The present work examines the effect of each one of these terms on the SR reconstruction result with respect to the presence or absence of noise in the low-resolution (LR) frames. Experimentation is carried out with the widely employed L₂, L₁, Huber and Lorentzian estimators for the data-fidelity term. The Tikhonov and Bilateral (B) Total Variation (TV) techniques are employed for the regularization term. The extracted conclusions can, in practice, help to select an effective SR method for a given sequence of LR frames. Thus, in case that the potential methods present common data-fidelity or regularization term, and frames are noiseless, the method which employs the most robust regularization or data-fidelity term should be used. Otherwise, experimental conclusions regarding performance ranking vary with the presence of noise in frames, the noise model as well as the difference in robustness of efficiency between the rival terms. Estimators employed for the data-fidelity term or regularizations stand for the rival terms.