CUDA-Powered EDSR x4: Super-Resolution for Real-Time Video Enhancement

Abstract

Super-Resolution enhances image or video quality by upscaling frames to higher resolutions, which is essential for applications like investigative analysis demanding higher quality. However, this process is resource-intensive. GPUs, with their thousands of CUDA cores, offer significant parallel processing advantages over CPUs, enabling faster performance. This paper presents an optimized approach for real-time video enhancement using the EDSR x4 model with Nvidia-CuDNN acceleration. We employ dynamic CPU-GPU load balancing to distribute computational tasks based on resource availability, reducing processing time by 18% and achieving real-time upscaling with a processing time of 205 ms per 10frames. EDSR, originally designed for Single Image Super-Resolution (SISR), is chosen over Video Super-Resolution (VSR) methods due to its superior frame-level clarity, making it ideal for scenarios where individual frame quality is critical. A notable discovery is the wave-like behavior in normalized PSNR, SSIM, and VIF metrics across resolutions, revealing aperiodic relationship between resolution and perceived quality. This insight further informs optimal resolution selection for various applications. The proposed system efficiently handles 480p to 4K video, maintaining high image quality and GPU utilization between 60%-80%, making it suitable for real-time applications that require both speed and high fidelity.