ABSTRACT This paper proposes an efficient SRGAN‐based super‐resolution framework for VCD video enhancement, capable of producing high‐quality upscaled images with significantly reduced computational complexity. To achieve this goal, we replace the residual‐in‐residual dense block (RRDB) used in Real‐ESRGAN with a novel residual‐in‐residual sparse block (RRSB) and further apply similarity‐based pruning techniques to RRSB for lightweight optimization. Additionally, we introduce the Improved residual‐in‐residual sparse block (IRRSB), which reduces both the number of input variables and the number of modules within each block. Our approach achieves an 85% reduction in parameters and a 79% decrease in computational workload compared to the original architecture. The framework is specifically designed to upscale old film images, such as those from VCD sources, to HDTV resolution by processing low‐resolution inputs and generating outputs with up to 9–16× pixel magnification while effectively minimizing artifacts and blurring. Objective evaluations utilize standard metrics including peak signal‐to‐noise ratio (PSNR), structural similarity index (SSIM), and natural image quality evaluator (NIQE). Despite the significant reduction in computation, PSNR decreased by only 0.7% dB and SSIM by 3%, while NIQE improved by 13%, indicating an overall enhancement in natural image quality. These results demonstrate that the proposed IRRSB framework maintains strong performance while significantly reducing model size and computational complexity.
Hsia et al. (Thu,) studied this question.