Fast, Efficient and Lightweight Compressed Image Super-Resolution Network for Edge Devices

초록

In many applications, images are reduced in size and compressed to save storage and transmission bandwidth. This process leads to loss of detail and often generates undesirable artifacts that degrade visual quality and impact the performance of vision tasks. To solve this challenge, many studies have been proposed on compressed image superresolution (CISR). However, most previous works have designed complicate architectures that require substantial computational resources, limiting their applicability in edge devices. To address this problem, we propose a fast, efficient and lightweight compressed image super-resolution network (FELCSRN) for edge devices. The proposed FELCSRN is a single network that reduces compression artifacts and enhances the resolution simultaneously. Furthermore, the reparameterization and quantization methods are utilized to further reduce computational and memory costs. Experimental results demonstrate that the proposed FELCSRN outperforms existing efficient super-resolution methods in terms of quality metrics and efficiency. In addition, compared to state-of-theart CISR methods, it significantly reduces computational costs and model size. As a result of evaluating the performance of the proposed FELCSRN by deploying it on the Xilinx ZCU104 board, it was confirmed that CISR tasks are performed in realtime.