Super-resolution techniques for high-resolution image reconstruction
Date of Issue2014
School of Electrical and Electronic Engineering
This thesis investigates a number of techniques and algorithms for super resolution (SR) image reconstruction. In digital imaging applications, high resolution (HR) images with more details are often desired for human interpretation or machine learning of the image contents. In many commercial digital cameras, however, the quality of the captured images may be limited by many real-life constraints, including aliasing effects, lens or motion blurring, color filter and shot noise. Compared with upgrading hardware, performing software post-processing on the captured degraded low resolution (LR) images is an economical and effective alternative solution to obtain a HR image. SR is such an image processing technique which can reconstruct a HR image from one or a sequence of LR counterparts. In the first part, several algorithms are proposed to handle multi-frame HR image reconstruction from a sequence of LR images which contain an overlapping content of the same scene. The basic principle of multi-frame SR is to combine the unique information exploited from each of the LR images to obtain a HR image. Therefore, the two main issues for successful multi-frame SR reconstruction are exploring the motion among the LR images followed by fusing the LR images based on the estimated motion parameters. Outliers occurred in the LR images may decrease the accuracy of image registration and subsequently introduce artifacts to the reconstructed HR image. In Chapter 3, we propose a L1-norm approach to handle SR reconstruction from multiple LR images that may contain outliers. The adopted L1-norm technique is robust to the registration error and outliers existing in the captured LR images. Further, image registration and SR are combined into a single process and they are performed simultaneously. An iterative scheme ensures the improved estimates of HR image and motion parameters are obtained progressively. Another fundamental problem for multi-frame SR is the selection of a specific motion model to describe the relationship of the LR images. Most of the existing SR approaches assume the relative motion between LR images to be translational or sometimes rotational. This in-plane motion model may not be applicable in such SR applications when relative zooming motion exists among the captured LR images. In Chapter 4, we develop a joint image registration and SR framework for LR images with zooming motion. A general motion model consisting of translation, rotation and zooming is adopted to handle more real-life applications. The proposed method performs image registration and zooming SR simultaneously, instead of regarding them as two disjoint steps. Further, the proposed method develops an adaptive weighting scheme in the HR reconstruction. The second part of this thesis focuses on single frame SR problems. Different from multi-frame SR, less information about the scene is provided for single frame SR. Therefore, strong prior information of the target image is required in order to obtain a stable solution. In Chapter 5, we propose a reconstruction-based single frame vehicle license plate SR approach using learning patches. The example patches are selected from a specific training database which consists of example images sharing similar characteristic with the target image. A Maximum a posteriori (MAP) framework is adopted and the example patches are incorporated into the cost function as prior information to regularize the final solution. Different from the conventional hard-binary learning based methods which use a threshold to select the patch or not, the proposed method employs a soft weighting strategy based on the character prior information of license plate. The underlying idea is that patches from the categories with higher matching scores will be given higher weightings, and verse vice. In the third part, we deal with color SR reconstruction problems. It is a challenging task to reconstruct a HR color image from a sequence of LR mosaiced color images. Conventional approaches usually can be considered as a two-stage process of color demosaicing followed by color image SR. However, these approaches tend to produce suboptimal solutions as the color artifacts arising from the color demosaicing process may be propagated to the subsequent HR reconstruction. In Chapter 6, we propose a joint framework to combine color demosaicing and color SR into a single process. To suppress the color artifacts, a color correlation term is adopted to ensure consistency amongst all the color channels. An edge-directed total variation (TV) regularization technique is developed to preserve the high-frequency details of the desired HR image, especially in the regions consisting of edges. The edge information is explored based on the reconstructed HR image in the previous iteration. An iterative scheme ensures the improved HR color image is obtained.
DRNTU::Engineering::Electrical and electronic engineering::Electronic systems::Signal processing