【作者】 陈湘骥；

【导师】 韩国强；

【作者基本信息】 华南理工大学 ， 计算机应用技术， 2014， 博士

【摘要】 图像超分辨率重建是指从一幅或相关的多幅低分辨率图像中重建出高分辨率图像的过程。图像超分辨率重建技术最初在航天、气象、医学等领域得到研究和应用，随着数字时代的到来，这项技术走向更广阔的应用领域，有着巨大的应用前景和社会价值。同时，由于图像超分辨率重建是一个不适定的逆问题，而且涉及的技术领域广，因而该课题目前仍然面临着诸多挑战，是图像处理技术研究的热点领域。本文探讨了图像超分辨率重建的相关理论和方法，对图像超分辨率重建的各个环节进行了系统地、深入地研究，并重点研究了基于多尺度相似学习在图像超分辨率重建中的应用，以及相关高性能算法的设计。本论文研究的方向包括：1.多尺度相似学习在单图像超分辨率重建中的应用；2.多尺度相似学习在多图像超分辨率重建中的应用；3.高性能的图像和视频超分辨率重建算法的设计。针对上述问题，本文的主要研究内容及贡献归纳如下：1.比较了基于插值、基于正则化、基于多图像重构和基于学习的超分辨率重建算法，并在此基础上重点研究了基于学习的超分辨率重建算法，包括基于流形学习的近邻嵌入超分辨率重建算法和基于稀疏表示的超分辨率重建算法。阐明了多尺度自相似学习既可以直接为图像重建提供信息，也可以作为图像重建过程中的正则化条件的思路，并由此进一步提出了基于图例学习和多尺度自相似学习相结合的单图像超分辨率重建算法。算法提供了一个框架，通过建立联合字典的方式，使图例学习获得的先验和自身学习获得的信息共同发挥作用。同时考虑到，在图像的多尺度空间，图块间的相似关系具有传递性，因而引入了迭代算法。重建的图像边界锐利，同时对噪声有较好的抑制。2.研究了超分辨率放大倍数和重建图像的质量之间的关系，在此基础上，提出了基于稀疏表示的图像多级超分辨率重建算法。通过建立多尺度图像空间中的超分辨率重建模型，将单级的图像超分辨率重建转化为多级的重建，并为每一级图像重建训练专属字典，重建的图像在相同放大倍数情况下比单级重建的图像质量明显提高。3.研究了图像序列中的相似信息，阐明了如果同一物体或相似物体以相同或不同的尺度出现在多幅图像当中，则为利用存在的互补信息，实现图像的超分辨率重建提供了可能，据此提出了基于多图像相似学习的超分辨率重建算法。算法针对传统的多图像超分辨率重建方法对图像配准要求高、重建质量不稳定的问题，提出采用图像拆分和多重配准的策略，以提升图像配准的精度；同时将基于学习的近邻嵌入方法引入多图像超分辨率重建当中，以降低像素填充环节对图像配准的要求。一升一降使算法具有较好的鲁棒性，增强了实用价值。4.对基于学习的超分辨率重建技术在视频中的应用进行了探索。针对视频应用实时性要求高的特点，提出了基于GPU加速和稀疏表示的视频超分辨率重建算法。着重于研究算法的并行优化、数据流的优化和资源使用的优化三个方面，通过设置视频帧队列、提高显存访问并发率、设计和采用了高效并行的主成分分析降维和相似度计算等手段，使算法执行速度比现有CPU串行算法提高了2个数量级。更多还原

【Abstract】 The image super-resolution reconstruction refers to the process that reconstruct high-definition image from one or several low resolution images. The technology of super-resolution reconstruction was initially studied and applied in aerospace, meteorology,medicine and other neighborhood, but as the digital times is coming, it is moving towards themass, and gradually heading for greater application areas, therefore, it has great applicationprospects and society and economy value. Simultaneously, the topic of the technology ofsuper-resolution reconstruction still faces much more challenges at present, for it is an ill-posed problem and its involving technologies are wide, which is the major field in theresearch of image processing technology.This paper discusses the relative theories and methods of the image super-resolutionreconstruction, systematically and thoroughly studies all aspects of the image super-resolutionreconstruction, focuses on the image reconstruction in super-resolution based on the multi-scale similarity learning, and designs the relative high-performance algorithms. The researchdirections of this paper include:1. the application of the multi-scale similarity learning in thesingle-image super-resolution reconstruction;2. the application of the multi-scale similaritylearning in the multi-image super-resolution reconstruction;3. the design of super-resolutionreconstruction algorithm for high-performance images and video. In view of the aboveproblems, this paper sums up the main research contents and contributions as follows:1. It compares the super-resolution reconstruction algorithms based on the interpolation,regularization, multi-image reconstruction and learning, and on this basis, it focuses on thesuper-resolution reconstruction algorithm based on learning, including the neighborembedding algorithm based on the manifold learning and the super-resolution reconstructionalgorithm based on the sparse representation. In addition, it illustrates the multi-scale self-similarity learning can not only provide information directly to the image reconstruction, butit also can act as the idea of regularization in image reconstruction, and thus it puts forwardthe single image super-resolution reconstruction algorithm based on a combination of legendlearning and multi-scale self-similarity learning. The algorithm provides a framework, whichmake the priori gained from the legend learning and the information obtained from the selflearning work together by establishing a joint dictionary. Meanwhile, it introduces an iterativealgorithm, since considering that the similarity relation among blocks of graph istransmissible in the multi-scale space of image. The reconstructed image boundary is sharp, which can effectively suppress noise.2. It studies the relationship between the super-resolution magnification and thereconstructed image quality, and on this basis, it puts forward the multi-image super-resolution reconstruction algorithm based on sparse representation. The single-image super-resolution reconstruction is transferred into multi-image reconstruction by building the super-resolution reconstruction in the multi-scale image space, and compared to the singlereconstructed image; the reconstructed image quality by using an exclusive dictionary foreach level of the image reconstruction shows a significant improvement at the samemagnification conditions.3. It studies the similar information in the image sequence, illustrates that an existingcomplementary information shall be adopted if the same object or similar objects appear inmultiple images with the same or different scales, which provides the possibility of realizationof image super-resolution reconstruction, and on this basis, it puts forward the super-resolution reconstruction algorithm based on the multi-image similarity learning. Against thetraditional multi-image super-resolution reconstruction method, the algorithm improves theprecision of image registration by using the image splitting and multiple registration strategyto solve the problems of high-demanding image registration and unstable reconstructedquality; simultaneously, the neighbor embedding method based on learning is introduced inthe multi-image super-resolution reconstruction, so as to reduce the pixel filling link’sdemand for the image registration. The rise and fall make the algorithm not only own goodrobustness, but also improve the practical value.4. It explores the application of the super-resolution reconstruction technology based onthe learning in the video. Aimed at the high-demanding real-time video application, the super-resolution reconstruction algorithm of video based on GPU acceleration and sparserepresentation is putted forward. It focuses on the parallel optimization of the algorithm, theoptimization of data stream and resource utilization. In addition, compared to the existingCPU serial algorithm, the execution speed of the algorithm improves significantly by2ordersof magnitude by setting the video frame queue, improving the memory access concurrency,and designing and adopting the efficient parallel Principal Component Analysis fordimensionality reduction and similarity computing and other methods.更多还原