Development of Efficient Methodologies for Image Super-Resolution Reconstruction

Abstract

High resolution (HR) images contain more image details, offer better visual perception and hence successfully applied in various imaging applications. The resolution of a captured image exclusively depends on the properties and
specification of the capturing device. Achieving an optimal balance between the enhancement of image resolution and the computational cost by redesigning the hardware is critically challenging. Super-resolution reconstruction (SRR) has
been established as an emerging off-line technology which provides a low cost and efficient software level solution to enhance the resolution of the captured low resolution (LR) images. This thesis is focused on development of robust and
computationally efficient SRR methodologies by addressing the problems related to reduction of noise, removal of blur and enhancement of spatial resolution of the LR image under two different frameworks, i.e., multi-image SRR and single-image SRR. Iterative back-projection-based, regularization-based SRR methodologies are proposed under multi-image SRR and learning-based methodologies are proposed under single-image SRR framework.
IBP is a popular, computationally cheap and straight forward multi-image SRR approach but suffers from local minima trapping, slow rate of convergence, and prone to ringing and jaggy artifacts. To overcome these pitfalls, enhanced
IBP based approaches are proposed by incorporating better initial guess, efficient back-projection kernels and metaheuristic algorithms to minimize the total reconstruction error accumulated during the iterative process.
SRR of image is an ill-posed inverse problem and needs to be regularized to obtain a unique and stable solution. Here, the proposed regularization-based method imposed efficient regularization constraints which maintain consistency
of structural features and enable to preserve the complete set of high frequency components lost in the image acquisition process. The utilized regularization constraints are based on the local phase coherence (LPC) of the image. LPC primarily depends on the phase information hence it is highly immune to degradation, and invariant to the local contrast. Moreover, it is insensitive to the intensity variation and noise incurred in the image. Consequently, this method is robust to illumination and/or intensity variations in images and favors a well-behaved stable solution with suppressed artifacts.
Resolution enhancement via multi-image SRR methodologies solely depends on the accuracy of the estimation of registration parameters of the LR observations. To address this issue, a SRR method is proposed which exploits the weighted averaging of the neighborhood pixels to estimate each pixel of the output HR image without performing exact image registration, de-blurring and de-noising of available LR images. The weight in the reconstruction process exhibits the correlation between pixels and combined Pseudo-Zernike moment invariants (CPZMIs) of image pixels are utilized to measure these correlations.
In real-time applications, the LR images undergo different local or complex motions among each other. In such type of situation, the performance of the above method degrades. Moreover availability of a series of LR observations of a
particular scene encompassing sub-pixel shift among each other is not an easy task always. As an alternative, learning-based single-image SRR (LSISRR)methodologies are proposed here.
LSISRR methods utilize numerous amounts of external training image dataset to provide visually appealing HR image from a single LR image at a cost of high computational overload. To overcome these shortcomings, two methods are proposed which utilize automated schemes for the selection of best training
image dataset and integrates faster searching schemes for the selection of candidate patches. The first LSISRR method is based on the framework of Markov Random Field (LSISRR-MRF) approximated by belief propagation (BP) algorithm followed by the proposition of a novel LSISRR method namely SRR via rigid body molecular docking (RiBMD-SR) based on central idea of Molecular docking (MD) to estimate the HR image. MD is an essential tool in the drug designing process which helps in predicting the most appropriate configuration as well as the optimal interaction energy between the interacting molecules such as, ligand and protein to form a stable complex. Similarly, in RiBMD-SR method estimation of the HR patch is achieved by optimizing the interaction energy between the observed LR patch and the appropriately chosen candidate patches from the training database. The LSISRR-MRF method provides a high-quality solution at a cost of more computational overload whereas, the RiBMD-SR method provides an efficient trade-off between computational speed and image-quality.
All the proposed reconstruction methods are validated with simulated as well as real-time test images. Performance evaluations of these methods are compared with their related state-of-the-art methods both qualitatively and quantitatively.