A Novel Video Super-Resolution Enhancement Method Based on Residual Learning Using Hidden Markov Random Fields and a New Deep Learning Network Architecture

In today's world, improving the quality and clarity of videos has become increasingly important, particularly in the fields of surveillance, medicine, and imaging technologies. Traditional super-resolution methods primarily focus on the full reconstruction of video frames, which poses challenges in preserving fine details and complex structures. This paper introduces a novel approach based on parallel deep networks, effectively enhancing video quality by dividing video frames into three separate input branches: raw images, outputs based on Hidden Markov Random Fields (HMRF), and temporal images. The method also leverages techniques such as residual learning and random patching within a unified framework that combines spatial segmentation (HMRF) and temporal information. This integration allows the model to better capture spatial and temporal dependencies, leading to more accurate and efficient video frame reconstruction. To better focus on high-frequency details and mitigate the vanishing gradient problem, residual learning is employed, enabling the network to estimate only the additional details necessary for reconstructing high-resolution images. Additionally, through random patching, the network training process is designed to emphasize critical features and intricate textures. Experimental results demonstrate that the proposed method achieves an SSIM of 0.92857 and a PSNR of 34.8617, offering superior clarity in video reconstruction.