Recent advances in display technology have led to the widespread use of 1080p (FHD) and 4K (UHD) displays, increasing the importance of single-image super-resolution (SISR), a technique that uses low-resolution images to create high-resolution images. Existing research on SISR has focused on improving the accuracy in constructing deep learning networks with significant computational complexity. However, since devices such as TVs, VR headsets, and autonomous systems have limited memory, it is necessary to research real-time image super-resolution.
In this paper, we propose CASR, an image super-resolution network that transforms compressed low-resolution images into 4K images in real-time. We propose a cascade upsampling with channel alignment method for upsampling, and reparameterization convolution layer to improve the performance of the network without any additional computational cost. In addition, we used the distillation method to improve performance and time.
Experimental results show that the proposed network achieves 22.80 dB PSNR and 0.56 ms inference time on RTX 3090 devices when evaluated on the AIS 2024 real-time super-resolution validation scale X4 dataset. The code is available at https://github.com/rlghksdbs/CASR
Top : CASR train stage.
Bottom : CASR test stage.
This work was supported by The KEIT(Korea planning \& Evaluation Institute of Industrial Technology) grant funded by the Korea government(MOTIE) (No.20014107, Development of 3D camera technology with fusion of thermal images to cope with the night and day environment)
@InProceedings{Yoon_2024_CVPR,
author = {Yoon, Kihwan and Gankhuyag, Ganzorig and Park, Jinman and Son, Haengseon and Min, Kyoungwon},
title = {CASR: Efficient Cascade Network Structure with Channel Aligned method for 4K Real-Time Single Image Super-Resolution},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2024},
pages = {7911-7920}
}