Ordinal Regression Based Subpixel Shift Estimation for Video Super-Resolution
- PDF / 1,780,393 Bytes
- 9 Pages / 600.03 x 792 pts Page_size
- 91 Downloads / 188 Views
Research Article Ordinal Regression Based Subpixel Shift Estimation for Video Super-Resolution Mithun Das Gupta,1 Shyamsundar Rajaram,1 Thomas S. Huang,1 and Nemanja Petrovic2 1 Department 2 Google
of Electrical and Computer Engineering, University of Illinois, Urbana Champaign, IL 61801-2918, USA Inc., 1440 Broadway, New York, NY 10018, USA
Received 2 October 2006; Accepted 3 May 2007 Recommended by Richard R. Schultz We present a supervised learning-based approach for subpixel motion estimation which is then used to perform video superresolution. The novelty of this work is the formulation of the problem of subpixel motion estimation in a ranking framework. The ranking formulation is a variant of classification and regression formulation, in which the ordering present in class labels namely, the shift between patches is explicitly taken into account. Finally, we demonstrate the applicability of our approach on superresolving synthetically generated images with global subpixel shifts and enhancing real video frames by accounting for both local integer and subpixel shifts. Copyright © 2007 Mithun Das Gupta et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
1.
INTRODUCTION
Shift estimation between two or more frames from a video has been of constant interest to researchers in computer vision. Need for accurate shift estimation arises from many practical situations. Applications, such as video frame registration, resolution enhancement, super-resolution, and optical-flow-based tracking, depend on reliable techniques for shift estimation for accuracy. Consequently, the accuracy of shift estimation methods is of utmost importance for these applications. Since the Lucas-Kanade [1] algorithm was proposed in 1981, image alignment has become one of the most significant contributions of computer vision. Applications of Lucas-Kanade image-alignment technique range from optical flow, tracking, layered motion to mosaic construction, medical image registration, and face coding. The principal idea of their technique was the introduction of imagegradients to infer the location of the target image patch in the subsequent frames which were best matches based on some arbitrary similarity metric. Many researchers have come up with refinements of their technique, to compute gradients in smarter ways, or to select the search region in smarter ways, but the principal idea has remained the same. Detailed reviews on motion estimation has been done by Aggarwal and Nandhakumar [2], Mitiche and Bouthemy [3], and Nagel [4]. Three main approaches to motion estimation can be
identified as estimation based on spatial-gradient, image correlation, and regularization of spatiotemporal energy. A closely related problem which has not yet received much focus in the literature is the problem of subpixel shift estimation which is harder than estimating shifts with pixel accuracy.
Data Loading...
