Comparative analysis of image super-resolution: A concurrent study of RGB and depth images

The introduction of deep learning has significantly advanced super-resolution techniques, allowing for the extraction of detailed textures and nuances vital for accurately analyzing visual information. The main contribution of our study is the development of the iDFD-SR dataset, an enhancement of the existing Indoor Depth from Defocus (iDFD) dataset, designed to evaluate super-resolution algorithms in the complex and varied indoor settings. Utilizing cutting-edge deep learning models such as EDSR (Enhanced Deep Super-Resolution Network), ESRGAN (Enhanced Super-Resolution Generative Adversarial Networks), RDN (Residual Dense Network), LIIF (Local Implicit Image Function) and HAT (Hybrid Attention Transformer), we perform a thorough comparison across several datasets—All in Focus (AIF), Out of Focus (OOF), Raw Depth (DR), and Depth after In-Painting (DI)—to identify the strengths and challenges of each approach. Our findings demonstrate notable enhancements in image quality for both RGB and depth images. However, the research acknowledges limitations, notably its indoor setting focus and the extensive computational resources required. Future work aims to broaden the application of super-resolution techniques, increase computational efficiency, and enhance the accuracy of depth estimations.