1 of 28

Selective Residual M-Net for Real Image Denoising

Β 

Β 

Β 

2 of 28

Outline

Introduction

01

Related Work

02

Method

04

Conclusion

05

03

Experiment

3 of 28

01

Introduction

Method

02

Experiment

04

Conclusion

05

Related Work

03

4 of 28

Introduction

01

4

(1/2)

Β 

Β 

Β 

Forward problem

Β 

ill-posed problem

Β 

5 of 28

Introduction

01

5

(2/2)

  • Contribution
  • We propose the improved hierarchical architecture (M-Net+) model SRMNet for both synthesized additive white Gaussian noise (AWGN) and real-world noise.

​

  • We propose the efficient extraction block called selective residual block (SRB) which is improved from the residual dense block (RDB) for image super-resolution.

​

  • We experiment on 2 synthesized image datasets and 2 real-world noisy datasets to demonstrate that our proposed model achieves the state-of-the-art in image denoising quantitatively and qualitatively with less computational complexity.

6 of 28

Outline

Introduction

01

Method

03

Experiment

04

Conclusion

05

Related Work

02

7 of 28

Introduction

01

Related Work

Experiment

04

Conclusion

05

02

Method

03

8 of 28

02

Related Work

8

(1/2)

  • M-Net

Mehta, Raghav, and Jayanthi Sivaswamy. "M-net: A convolutional neural network for deep brain structure segmentation."Β 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017). IEEE, 2017.

9 of 28

02

Related Work

9

(2/2)

  • Selective Kernel Network (Feature Fusion)

Zamir, Syed Waqas, et al. "Learning enriched features for real image restoration and enhancement."Β European Conference on Computer Vision. Springer, Cham, 2020.

Li, Xiang, et al. "Selective kernel networks."Β Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.

10 of 28

Outline

Introduction

01

Method

03

Experiment

04

Conclusion

05

Related Work

02

11 of 28

03

Method

Introduction

01

Related Work

02

Conclusion

05

Experiment

04

12 of 28

03

Method

12

(1/6)

  • Selective Residual M-Net (SRMNet)

13 of 28

03

Method

13

(2/6)

  • Selective Residual M-Net (SRMNet)
  • Selective Residual Block (SRB)
    • Improved from Residual Dense Block (RDB)
    • Low computation complexity

​

  • Resizing Module
    • Main U-Net model architecture
    • Gatepost feature path

​

  • Improved hierarchical M-Net+
    • Main U-Net model architecture
    • Encoder gatepost feature path
    • Decoder gatepost feature path

​

簑稱

14 of 28

03

Method

14

(3/6)

  • Selective Residual Block (SRB)

conv

conv

ReLU

conv

ReLU

conv

ReLU

1 x 1

conv

  • RDB

15 of 28

03

Method

15

(4/6)

  • Resizing module
  • Pixel (Un)Shuffle up- (down-) sampling

16 of 28

03

Method

16

(5/6)

  • Improved hierarchical M-Net+
  • Main U-Net model architecture
    • Extraction block (original: conv-BN-ReLU)
    • Resizing modules (original: max-pooling)

​

  • Encoder gatepost feature path
    • Additional shallow feature extraction (original : directly concat)

​

  • Decoder gatepost feature path
    • Selective Kernel Feature Fusion (SKFF) (original: directly concat)

​

17 of 28

03

Method

17

(6/6)

  • Other details

18 of 28

Outline

Introduction

01

Method

03

Experiment

04

Conclusion

05

Related Work

02

19 of 28

04

Experiment

Introduction

01

Related Work

02

Conclusion

05

Method

03

20 of 28

Experiment

04

(1/5)

20

  • Dataset
  • Real-world noise
    • Train SIDD train set
    • Validation: SIDD benchmark
    • Test: SIDD, DND

​

​

​

  • Synthesized additive white gaussian noise (AWGN)
    • Train: DIV2K train set
    • Validation: DIV2K val set
    • Test: CBSD68, Kodak24

21 of 28

21

Experiment

04

(2/5)

  • Real-world denoising results

22 of 28

22

Experiment

04

(3/5)

  • Synthesized gaussian denoising results

23 of 28

23

Experiment

04

(4/5)

  • Visual performances (CBSD68, SIDD)

24 of 28

24

Experiment

04

(5/5)

  • Efficiency and model complexity

25 of 28

Outline

Introduction

01

Method

03

Experiment

04

Conclusion

05

Related Work

02

26 of 28

05

Conclusion

Introduction

01

Related Work

02

Experiment

04

Method

03

27 of 28

27

Conclusion

05

(1/1)

  • Conclusion
  • We propose the improved hierarchical architecture (M-Net+) model SRMNet for both synthesized additive white Gaussian noise (AWGN) and real-world noise.

​

  • We propose the extraction block called selective residual block (SRB) which is an efficient block compared with the RDB.

​

  • Our proposed model achieves the state-of-the-art performances on image denoising with less computational complexity.

28 of 28

Thanks for Listening

Speaker：Chi-Mao Fan

Date : 2022.08.30