1 of 29

Half Wavelet Attention on M-Net+ for

Low-light Image Enhancement

Β 

Β 

Β 

2 of 29

Outline

Introduction

01

Related Work

02

Proposed Method

04

Conclusion

05

03

Experiment

3 of 29

01

Introduction

Proposed Method

02

Experiment

04

Conclusion

05

Related Work

03

4 of 29

Introduction

01

4

(1/2)

Low-Light Image Enhancement is a low-level computer vision task which intensifies the dark images to appropriate brightness. It can also be seen as an ill-posed problem in image restoration domain.

Β 

Β 

Image Enhancement

5 of 29

Introduction

01

5

(2/2)

  • Contribution
  • We propose the improved hierarchical backbone architecture (MNet+) for the task of low-light image enhancement.

​

  • We propose the efficient feature extraction block called half wavelet attention block (HWAB) which can gain more diversified features in another domain.

​

  • We experiment on two low-light image enhancement datasets to demonstrate that our proposed model achieves state-of-the-art performance in image enhancement in terms of quantity and quality and even with less computational complexity.

6 of 29

Outline

Introduction

01

Proposed Method

03

Experiment

04

Conclusion

05

Related Work

02

7 of 29

Introduction

01

Related Work

Experiment

04

Conclusion

05

02

Proposed Method

03

8 of 29

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 29

02

Related Work

9

(2/2)

  • Wavelet transformation

C. Chen and C. Hsiao, β€œHaar wavelet method for solving lumped and distributed-parameter systems,” IEE Proceedings-Control Theory and Applications, vol. 144, no. 1, pp. 87–94, 1997.

10 of 29

Outline

Introduction

01

Proposed Method

03

Experiment

04

Conclusion

05

Related Work

02

11 of 29

03

Proposed Method

Introduction

01

Related Work

02

Conclusion

05

Experiment

04

12 of 29

03

Proposed Method

12

(1/6)

  • Half Wavelet attention M-Net (HWMNet)

13 of 29

03

13

  • Half Wavelet attention M-Net (HWMNet)
  • Half Wavelet Attention block (HWAB)
    • 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

Proposed Method

(2/6)

14 of 29

03

14

  • Wavelet Attention block (HWAB)

Proposed Method

(3/6)

15 of 29

03

15

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

Proposed Method

(4/6)

16 of 29

03

16

  • 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)

​

Proposed Method

(5/6)

17 of 29

03

17

  • Other details

Proposed Method

(6/6)

18 of 29

Outline

Introduction

01

Proposed Method

03

Experiment

04

Conclusion

05

Related Work

02

19 of 29

04

Experiment

Introduction

01

Related Work

02

Conclusion

05

Proposed Method

03

20 of 29

Experiment

04

(1/6)

20

  • Dataset
  • Low-light image enhancement
    • LOL dataset
    • Training: 485
    • Testing: 15

​

  • Low-light image enhancement (image retouching)
    • MIT-Adobe FiveK dataset
    • Training: 4500
    • Test: 500

21 of 29

21

Experiment

04

(2/6)

  • Low-light image enhancement results (LOL dataset)

22 of 29

22

Experiment

04

(3/6)

  • Comparison of model complexity (FLOPs)

23 of 29

23

Experiment

04

(4/6)

  • Low-light image retouching results (MIT-Adobe FiveK dataset)

24 of 29

24

Experiment

04

(5/6)

  • Visual performances (LOL dataset)

25 of 29

25

Experiment

04

(6/6)

  • Visual performances (MIT-Adobe FiveK dataset)

26 of 29

Outline

Introduction

01

Proposed Method

03

Experiment

04

Conclusion

05

Related Work

02

27 of 29

05

Conclusion

Introduction

01

Related Work

02

Experiment

04

Proposed Method

03

28 of 29

28

Conclusion

05

(1/1)

  • Conclusion
  • We propose the improved hierarchical backbone architecture (M-Net+) model HWMNet for low-light image enhancement.

​

  • We propose the feature extraction block called half wavelet attention block (HWAB) focusing on the features of wavelet domain which can enrich the semantic information.

​

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

29 of 29

Thanks for Listening

Speaker：Chi-Mao Fan

Date : 2022.10.12