1 of 10

Lightweight Real-Time Image Super-Resolution Network for 4K Images����Ganzorig Gankhuyag*, Kihwan Yoon*, Jinman Park, Haeng Seon Son, Kyoungwon Min���* : Equal Contribution�����

2 of 10

Key Contribution

​

​

  • We proposed the lightweight real-time image super resolution (LRSRN) network structure that simultaneously achieves high accuracy and real-time speed.

​

    • Low computational complexity and high accuracy compared to traditional SISR methods

​

​

  • We employed a reparameterized convolution (RepConv) layer, which enhances image quality while maintaining model size and inference speed

1

3 of 10

Network Architecture

  • We applied RepConv to each convolution layer, which is a more efficient method

​

2

(a) The Training mode of proposed network

(b) The inference mode of proposed network

4 of 10

Reparameterized Block

  • We applied RepConv to each convolution layer, which is a more efficient method.
  • We applied an advanced version of RepConv block when In/Our channels are not equals

3

(a) In/Out channels are equals

(b) In/Out channels are not equals

(c) Inference mode

5 of 10

Experimental Results

4

Ablation study results on DIV2K val dataset

6 of 10

Experimental Results

5

Qualitative results comparison on benchmark datasets.

7 of 10

Experimental Results

6

Qualitative results comparison on NTIRE validation dataset

8 of 10

Experimental Results

7

9 of 10

Experimental Results

8

10 of 10

Paper and Code link

9

Code link

Paper link

gnzrg25@gmail.com

rlghksdbs@gmail.com

If you have any questions, please contact us