Super resolution imaging using modified lanr based on separable filtering

Recently, remarkable advances have been achieved in reconstructing high-resolution image from noisy, and low-resolution images. Reaching super resolution has been a challenge in image processing practices, because of their under-constrained nature that requires the missing HR image details to be...

Full description

Saved in:
Bibliographic Details
Main Author: Somadina, Ike Chidiebere
Format: Thesis
Language:English
Published: 2019
Subjects:
Online Access:http://psasir.upm.edu.my/id/eprint/77392/1/FK%202019%201%20ir.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Recently, remarkable advances have been achieved in reconstructing high-resolution image from noisy, and low-resolution images. Reaching super resolution has been a challenge in image processing practices, because of their under-constrained nature that requires the missing HR image details to be reconstructed. In this research, the long-established single-image super-resolution problem is addressed by integrating the multiresolution property of Wavelet and the flexibility of Locally Anchored Neighbourhood Regression model to formulate a novel edgebased single image super resolution algorithm that allows robust estimation of missing frequency details in wavelet domain with complete enhancement procedure. Firstly, the low resolution input image is decomposed into four frequency sub-bands, comprising of one approximate coefficient and three detailed coefficients sampled by applying discrete wavelet transformation. The underlying idea is to process and reconstruct information in low and high frequency sub-bands based on separable property of neighbourhood filtering to achieve fast parallel and vectorized operation, while enhancing algorithmic performance by reducing computational burden resulting from computing the weighted function of every pixel for each pixel in an image. We then processed the frequency sub-bands using the inverse discrete wavelet transforms which does not in any way increase image size, rather it reconstructs the original image with high integrity of preserved fine edge details and more realistic textures. Super resolution is then achieved using the regularized patch representation (projection matrix) learned to predict the high resolution image features. Lastly, we incorporate the nonlocal self-similarity prior to refine our reconstructed high resolution result; hence preserving the local singularity and edges details to achieve a more sophisticated, distinctive and robust image super resolution. Experimental results on standard images with qualitative and quantitative comparisons against several top-performing state- of-the-art SR methods demonstrate the effectiveness and stability of the proposed algorithm. The proposed method reaches the highest PSNR for scale factors of 2, 3 and 4, respectively for Set5 datasets with around 0.03- 0.70 dB better than LANR, and 0.2-1.60 dB better than the second best method, i.e. ANR. Similarly, we achieved around 0.03-1.10 dB better than LANR, and 0.2-1.80 dB better than ANR for scale factors of 2, 3 and 4 on Set14 dataset.