Super-Resolution via Image Warping
Terrance E. Boult, Ming‐Chao Chiang, Ross J. Micheals · Kluwer Academic Publishers eBooks · 2005
This chapter focuses on three issues: supporting image warping algorithms for super-resolution, examples of how image warping algorithms impact super-resolution image quality, and the development of quantitative techniques for super-resolution algorithm evaluation. The warping approach proposed in this chapter is based on the integrating resampler [Chiang and Boult, 1996] which warps the image while both enforcing the underlying image reconstruction and satisfying the imaging consistent constraint [Boult and Wolberg, 1993]. The imaging consistent constraint requires that the image reconstruction yields a function which, when convolved with the imaging system’s point-spread function (PSF), is consistent with the input image. Many popular reconstruction techniques, including bilinear and natural cubic splines, do not satisfy the imaging consistent constraint. In this chapter, we review imaging consistent warping algorithms, how they form the core of the integrating resampler, and their implementation. Although imaging consistent warping techniques can be used in other super-resolution implementations, such as those discussed in Chapter 8, we present its use in a simpler direct approach: warping followed by a straightfoward fusion. Examples are provided on grayscale images of simple patterns, text, and human faces. The use of priors in the fusion, such as those used in Chapter 10 could further enhance the results, but we analysize the simpler approach to isolate the impact of the warping algorithm. The chapter then discusses the important problem of quantitative evaluation and presents a summary of two different quantitative experiments: using OCR and face recognition as metrics. These experiments clearly show the importance of high-quality reconstruction and warping to super-resolution. Perhaps more importantly, these experiments show that even when images are qualitatively similar, quantitative differences appear in machine processing. As the super-resolution field is pushed towards its boundardies, the ability to measure progress, even if it is small, becomes increasingly important.