Imaging-consistent warping and super-resolution
Terrance E. Boult, Ming‐Chao Chiang · 1998
Aimed at establishing a framework for applications that require sufficiently accurate warped intensity values to perform their tasks, this thesis begins by introducing a new class of warping algorithms that provides an efficient method for image warping using the class of imaging-consistent reconstruction/restoration algorithms proposed by Boult and Wolberg. We show that by coupling the degradation model of the imaging system directly into the imaging-consistent warping algorithms, we can better approximate the warping characteristics of real sensors, and this, in turn, significantly improves the accuracy of the warped intensity values. We then present two new approaches for super-resolution imaging. This problem is chosen because it is not only useful in itself, but it also exemplifies the kinds of applications for which the imaging-consistent warping algorithms are designed. The image-based approach presumes that the images were taken under the same illumination conditions and uses the intensity information provided by the image sequence to construct the super-resolution image. However, when imaging from different view-points, over long temporal spans, or imaging scenes with moving 3D objects, the image intensities naturally vary. The edge-based approach, on edge models and a local blur estimate, circumvents the difficulties caused by lighting variations. We show that the super-resolution problem may be solved by direct methods, which are not only computationally cheaper, but they also give results comparable to or better than those using Irani's back-projection method. Our experiments also show that image warping techniques may have a strong impact on the quality of super-resolution imaging. Moreover, we also demonstrate that super-resolution provides added benefits even if the final sampling rate is exactly the same as the original. We also propose two new algorithms to address the issue of quantitatively measuring the advantages of super-resolution algorithms. The based approach uses OCR as the fundamental measure. The matching and pose estimation based approach uses appearance matching and pose estimation as the primary metric and image-quality metric as a secondary measure. We show that even when the images are qualitatively similar, quantitative differences appear in machine processing.