Abstract
A variant of the general super-resolution problem involves the generation of a single high-resolution image using a sequence of low-resolution video frames [1]. We study a related problem in which a high-resolution image is obtained from multiple low-resolution cameras. In our model each camera measurement represents a transformation of the object space according to some (a priori known) position/magnification/rotation. Note that an array of imagers can be characterized by their number as well as their range of positions/magnifications/rotations. We refer to the range of position/magnification/rotation as the measurement diversity of the array. Large diversity might correspond to a range of positions = 16 pixels, a range of rotations = 45 degrees, and a range of magnifications = 20 %. Medium diversity might be considered 8 pixels, 15 degrees, and 10%. Random noise is also included within our imager model. We have developed several algorithms for combining data from disparate camera measurements in order to arrive at a single high-resolution object estimate. The first algorithm that we investigated was a variant of the gradient-based Iterative Back-Projection (IBP) algorithm [2], extended to handle general affine transformations. The results of applying this algorithm are shown in figure 1, in which we see that increasing the number of cameras improves reconstruction mean-squared-error (MSE). This figure also demonstrates the important trend that (for a sufficient number of cameras) increased measurement diversity provides improved MSE performance.
© 2003 Optical Society of America
PDF ArticleMore Like This
Miguel A. Preciado, Guillem Carles, and Andrew Harvey
JTu5A.12 3D Image Acquisition and Display: Technology, Perception and Applications (3D) 2017
H. Kikuta, K. Numata, M. Muto, K. Iwata, H. Toyota, K. Moriwaki, T. Yotuya, and H. Sato
ThRR3 Frontiers in Optics (FiO) 2003
James Downing, Andrew A. Murray, and Andrew R. Harvey
JW1A.6 Computational Optical Sensing and Imaging (COSI) 2012