多分辨率相机及图像超分辨率技术研究

It is important to achieve high resolution imaging system by exploiting different techniques. Image super resolution is the technique generally exploited to overcome the limitations of optical imaging systems by using the image processing algorithms.For SR methods,however, there are three problems left unanswered: 1) Most of these existing methods fail to consider the local geometrical structure in the space of the training data. Hence, these methods may obtain ringing artifacts and blurring reconstructed results. 2) The general way to take the simulated LR image as the test image is not proper for image super resolution methods. Hence, it is difficult to evaluate the quality of the reconstructed image generated from SR methods. 3)Most IQA methods fail to formulate perceptual meaningful features to effectively express the distortion level and find a way to pool them into a single quality score. To address the aformentioned problems, the constribution of our project can be summarized into the following:1)To solve the first problem, this project proposes a sparse coding algorithm named double sparsity regularized manifold learning (DSRML) that explicitly considers the local manifold structure information in the data space. By incorporating the LLE based manifold learning as a regularizer into the traditional sparse coding objective function,the proposed method can provide sparse representation with local topology structure information and achieve better reconstruction for image SR. 2) To solve the second problem, this project also proposes a new design of a multi-resolution imaging camera, which is exploited to study the relationship of the different resolution images. By adopting three-path light splitting technology, we can ensure that the illuminance of image captured on three CCDs of multi-resolution imaging camera is equal. This means that the LR images and HR image can be collected from the same scene at the same time by emploiting the multi-resolution imaging camera. The designed multi-resolution imaging camera can provide us with real images of different resolutions, which builds a solid foundation for evaluating various algorithms and analyzing the images with different resolution.3) To solve the third problem, a two-step full reference IQA algorithm is proposed in this paper. In the first step, the sparse representation model is employed to extract the perceptual meaningful features for the IQA problem. In the second step, machine learning technique is employed to mimic the complex characteristics of human vision system which pools the local distortion measurements obtained into a single image quality score. More specifically, the kernel ridge regression (KRR) is adopted to learn the relationship between local quality measurements of the patches and the quality score of the whole image.

目前图像超分辨率技术存在着三个制约超分重建精度的问题：1)在重建过程中未能考虑原始数据空间的几何结构信息，从而在重建过程中可能会导致性能的下降和伪像的产生。2)在超分辨率算法中，用来做测试的模拟低分辨率图像跟相机所拍摄的同一场景的低分辨率真实图像是不一样的，这就给重建结果的质量评价带来很大困难。3)针对超分辨率重建结果，目前的质量评价方法未能提取有效的特征来表达每个图像块的局部质量分/失真程度，从而很难有效地进行质量评价。这三个问题是相互促进，相互影响。这是因为研究图像的超分重建算法、图像获取模式的选择与图像质量评价的关系将为图像质量改善提供强有力的理论依据。基于此，本文拟提出基于局部线性嵌入稀疏编码学习的超分辨率重建方法并设计多分辨率相机来探讨不同的采集模式对重建结果的影响。最后，我们在多分辨率相机所采集的真实图像的基础上，拟提出两阶段的图像质量评价模型对重建算法进行评价。

随着计算机的发展，提出一种有效的超分辨率重建算法来提高图像分辨率是图像处理领域追求的目标。基于学习的这种类型重建算法是目前超分重建算法研究的主流。然而这种类型算法普遍存在着以下两个问题：1)常规算法在学习过程中未能考虑原始数据空间的几何结构信息，从而可能会导致学习性能下降和重建过程中伪像的产生。2)由于在超分算法中用来做测试的模拟低分辨率图像跟相机所拍摄的同一场景的低分辨率真实图像是不一样的。这就给各种算法重建结果的质量评价带来了很大的困难。为了处理第一个问题，本文提出了基于局部线性嵌入稀疏编码学习的超分辨率重建方法。该方法通过考虑稀疏编码中的几何结构来改善超分辨率重建的质量。针对第二个问题，本文设计了多分辨率相机，该相机能在同一时刻对同一场景进行多分辨率成像。从而，可以利用多分辨率相机所采集的数据来验证各种重建算法的优劣以及通过分析不同分辨率图像之间的关系为导出更高分辨率的图像奠定基础。