Single Image Super-Resolution with the DIV2K Dataset: a case study

Iris Steenhout, Mahdi Asghari

Onderzoeksoutput: Working paper

32 Downloads (Pure)

Samenvatting

Single-image super-resolution (SISR) reconstructs high-resolution (HR) images from lowresolution (LR) inputs, addressing hardware limitations and enabling detailed visualization for diverse applications. This study explores the Efficient Sub-Pixel Convolutional Neural Network (ESPCN) for SISR tasks using the DIV2K dataset, focusing on computational efficiency and reconstruction quality. Unlike traditional methods relying on interpolation or iterative processes, ESPCN employs sub-pixel
convolution, providing an end-to-end, resource-efficient solution.

ESPCN demonstrates competitive performance, surpassing Bicubic, A+, and SRCNN in Peak Signal-to-Noise Ratio (PSNR) while maintaining a lightweight architecture suitable for real-time applications. However, its Structural Similarity Index (SSIM) suggests room for improvement in structural fidelity, and further evaluation of perceptual quality is needed due to the lack of comparative LPIPS scores.

This study highlights ESPCN’s ability to balance efficiency and performance, making it a practical choice for applications like medical imaging, surveillance, and autonomous systems. It contributes to advancing SISR by optimizing accuracy and resource use, paving the way for future developments in deep learning-based super-resolution.
Originele taal-2English
Pagina's1-8
Aantal pagina's8
StatusPublished - 7 jan 2025

Citeer dit