1 of 40

Image colorization & De-noise

Group 7

張凱博、高田、林芷萱、林彥承

2 of 40

outline

  • Introduction to Autoencoder
  • Image Colorization
  • Image De-noise
    • Native implementation
    • A Better Version
  • Comparison
  • Conclusion

3 of 40

Autoencoder

4 of 40

Autoencoder

5 of 40

https://speech.ee.ntu.edu.tw/~hylee/ml/ml2021-course-data/auto_v8.pdf from professor Hung-yi Lee

6 of 40

7 of 40

8 of 40

9 of 40

10 of 40

11 of 40

Image Colorization

12 of 40

What If We Input Grayscale Image?

13 of 40

Experiment

14 of 40

Training Dataset

92219 anime face images

​

15 of 40

Training Environment

Linux (ubuntu 18.04)

NVIDIA Tesla V100 SXM2 32GB * 8

Intel Xeon Platinum 8168 CPU 2.70GHz * 40

661G RAM available

16 of 40

Results

17 of 40

Training Set

18 of 40

Training Set

​

19 of 40

Training Set

​

20 of 40

Training Set

​

21 of 40

Testing Set Performance

​

22 of 40

Testing Set Performance

​

23 of 40

Testing Set Performance

​

24 of 40

Testing Set Performance

​

25 of 40

Testing Set Performance

​

26 of 40

Some Observations

27 of 40

Some Observations

​

28 of 40

Comparison (anime face)

convert with https://imagecolorizer.com/colorize.html

29 of 40

Comparison (anime face)

30 of 40

Image De-noise

31 of 40

32 of 40

Naive Implementation

We first conduct experiment on MNIST

33 of 40

34 of 40

35 of 40

Our first try on Anime Dataset

​

64

64

32

32

36 of 40

Our first try on Anime Dataset

Barely recognizable

37 of 40

Image De-noise

​

Add Noise

De-noised

38 of 40

Image De-noise

​

Add Noise

De-noised

39 of 40

Conclusion

  • The loss of testing set almost would not reduce after 50 epochs.
  • The hair and the eyes of some predicted anime images would have mixed color.
  • For different databsets image types, model would perform on different focus. So train the correspond type is important.

40 of 40

Q & A