Super-resolution image reconstruction from multiple low-resolution images

Abstract

This study explores a novel approach for super-resolution image reconstruction from multiple low-resolution images, employing frequency domain motion estimation technique (FMT), Keren-based image interpolation, and bicubic interpolation (BI). The method performs well in estimating scaling parameters, but accuracy decreases as shift distance or rotation angle increases. Compared to Vandewalle's algorithm, the proposed method shows better accuracy in estimating scaling parameters but similar accuracy for rotation and translation parameters. Differences are observed in estimated values for each parameter between both methods. The study underscores the need for further research to improve the accuracy of the proposed method in motion estimation and interpolation optimization. Additionally, Generative Adversarial Networks (GANs) outperform Bicubic and Wavelet Domain Super-Resolution (WDSR) algorithms in image quality improvement, indicated by higher Peak Signal-to-Noise Ratio (PSNR) values. This superior performance is attributed to GANs' ability to leverage deep learning algorithms to capture complex image features. The research validates the potential of the proposed method for super-resolution image reconstruction, and the power of deep learning-based algorithms, specifically GANs, in enhancing low-resolution images. More advanced motion estimation algorithms and interpolation technique optimization could further improve the accuracy of this method.