Enhanced Video Super-Resolution via a Dual-Stage Framework with VDSR and HMRF-DCNN

Video super-resolution (VSR) is a critical task in video processing, aiming to enhance the resolution of consecutive frames while maintaining visual quality. This paper presents a comprehensive approach to video super-resolution, integrating deep learning networks with Hidden Markov Random Field (HMRF) techniques. The proposed method consists of two stages: In the first stage, the Very Deep Super-Resolution (VDSR) technique is employed to sharpen frame edges and enhance resolution, prioritizing perceptually significant details by focusing on luminance components. Additionally, random patching optimizes VDSR performance by enhancing relevant image details while mitigating computational burdens. In the second stage, a parallel network integrates the output of the first phase, HMRF-based inputs, and chronological inputs to capture spatial and temporal dependencies for final resolution enhancement. This multi-faceted approach ensures superior resolution and visual quality in the final output frames. Experimental evaluation demonstrates significant improvement over existing methods, with a peak signal-to-noise ratio (PSNR) of 37.0295 and a structural similarity index (SSIM) of 0.94683. The proposed method presents a promising solution for high-quality video super-resolution, addressing the complex interplay of resolution enhancement and visual fidelity in video processing.