Zoom Based Image Super-Resolution: Using Two Level DWT as Feature Model

Abstract

Abstract

In this paper we present an algorithm of super-resolution (SR) imaging to reconstruct high-resolution (HR) image from sequence of low-resolution (LR) images of static scene captured at the different camera zoom factor. The resultant HR image is constructed at the resolution of the most zoomed LR image. In the proposed approach algorithm uses LR images of the static scene captured at three distinct zoom-factors. Learning based SR technique is used to enhance the spatial resolution of these LR images. For that training dataset comprising three sets of captured images which are LR images, enhances version of LR images-HR1 and enhanced version of HR1 images-HR2. Proposed algorithm use training images in two pairs of datasets LR-HR1 and LR-HR2 to obtain super-resolution of observed image. High-frequency details of the super-resolved image are learned in form of the two-level discrete wavelet transform coefficients of HR training images. The fine textures of the images are conserved with this approach. Finally, the super-resolved version of LR observations, captured at different zoom-factors, are combined. The suggested approach has been trailed on numerous real-world natural images. The results show that the suggested approach is better in both qualitative-SSIM and quantitative- PSNR, MSE manner over existing approaches.

Keywords: Discrete Wavelet Transform (DWT), Learning-based approach, Peak Signal to Noise Ratio (PSNR), Mean Square Error (MSE), Structural Similarity Index (SSIM)

Cite this Article

Meera D. Doshi, Prakash P. Gajjar, Ashish M. Kothari. Zoom Based Image Super-Resolution: Using Two-Level DWT as Feature Model. Journal of Communication Engineering & Systems. 2019; 9(2):
94–105p.