【信息技术】【2003】基于精确图像配准的超分辨率实现

本文为西澳大利亚大学（作者：Douglas Lim）的论文，共77页。

人眼几乎看不到的关键信息往往嵌入在同一场景拍摄的一系列低分辨率图像中。超分辨率使得能够通过重建单个图像来提取该信息，其分辨率高于任何单个图像中的分辨率。这在法医成像中特别有用，因为提取图像中的微小细节可以帮助解决犯罪问题。对同一场景拍摄的多个低分辨率图像的捕获会导致每个图像之间的失真。图像配准就是确定这种失真的过程，然后在超分辨率处理中使用该信息来创建一组模拟的低分辨率图像，再利用这些模拟图像和观测图像之间的差异来迭代更新高分辨率图像的初始估计。成功的超分辨率依赖于精确的图像配准。在本论文中，我们检验了当达到精确的图像配准时，重建高分辨率图像的视觉质量得到提高的假设。在本文的第一部分中，我们详细研究了图像的配准过程。图片和文本图像都使用两种算法进行了配准。第一种是基于优化方法的配准算法，另一种是基于RANSAC的算法。研究发现，仿射变换等高阶变换严重阻碍了优化算法的实现。这归因于需要优化的参数数量增加。在本文的第二部分，我们主要研究了超分辨过程。我们进行了许多实验来验证我们最初的假设。第一个实验涉及到在实现完全配准时重建图像，并将结果与使用RANSAC算法时的结果进行比较。结果表明，为了达到理想的配准效果，需要较高的重建图像视觉质量。我们还发现，与优化方法相比，使用RANSAC算法配准图像时，重建图像的视觉质量更高。

Crucial information barely visible to thehuman eye is often embedded in a series of low-resolution images taken of thesame scene. Super-resolution enables the extraction of this information byreconstructing a single image, at a higher resolution than is present in any ofthe individual images. This is particularly useful in forensic imaging, wherethe extraction of minute details in an image can help solve a crime. Thecapturing of multiple low-resolution images taken of the same scene results ina distortion between each image. Image registration is the process ofdetermining this distortion. This information is then used in thesuper-resolution process to create a set of simulated low-resolution images.The differences between these simulated images and the observed images are thenused to iteratively update an initial estimate of the high-resolution image.Successful superresolution is dependent on accurate image registration. In thisthesis, we examine the hypothesis that the visual quality of a reconstructedhigh-resolution image improves when accurate image registration is achieved. Inthe first part of this thesis, we examine the image registration process indetail. Both picture and text images are registered using two algorithms. Thefirst registration algorithm based on an optimization approach whilst the otheris based on the RANSAC algorithm. We find that the optimization approach isseverely hampered by higher degree transforms such as affine transforms. Thisis attributed to the increased number of parameters requiring optimizing. Inthe second part of this thesis, we focus on the super-resolution process.Numerous experiments were conducted to test our original hypothesis. The firstexperiment involved reconstructing an image when perfect registration wasachieved, and comparing the results to when the RANSAC algorithm was employed.The results suggested that the visual quality of the reconstructed images werehigher for perfect registration. We also found that the visual quality ofreconstructed images was higher when images were registered using the RANSACalgorithm, as compared to an optimization approach.

1. 引言
2. 图像配准
3. 图像配准算法
4. 配准算法的具体实现
5. 配准精度
6. 超分辨
7. 超分辨算法
8. 一种超分辨算法的具体实现
9. 重建图像的视觉质量
10. 结论
    附录A 配准合成图像的结果
    附录B 重建合成图像的结果

更多精彩文章请关注公众号：