Skip to content

somniumcaeruleumKonkuk/opensw23-LSJ

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

44 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Hello World!

Chapter

  • Team Introduction
  • Topic Introduction
  • Results
  • Analysis/Visualization
  • Installation
  • Presentation

Team Introduction

  • Name: 이승준(202211337)
  • Github ID: somniumcaeruleumKonkuk
  • Team type: Individual

Topic Introduction

Slide3.png Slide4.png Slide5.png Slide6.png

Results

Results

Analysis/Visualization

Original Image

original5.jpg

Noised Images

Fig2.png

Denoised Images

Fig3.png

Conclusion

  • Noise의 양(amount)과 강도(strength)가 클수록 디노이징(denoising) 잘 되지 않는 것을 볼 수 있었다. Noise의 강도(strength)가 작을 때는 양(amount)이 크게 영향을 주지 못했으나, Noise의 강도(strength)가 60을 넘은 후 부터는 양(amount)가 상당히 denoised image들의 화질에 영향을 미치는 중요한 변수임을 볼 수 있었다.
  • As the amount and strength of noise increased, you can see that the quality of the denoised images isn't good in the above experiment. When the strength was low, the amount of noise didn't matter that much. As the strength is over 60, however, the amount of noise was important variable to the quality.

Installation

  • Install numpy, opencv, tensorflow.

Environment

  • Python version: 3.10.10
  • Tensorflow version: 2.12.0
  • OS: Windows 11 Education (the latest version in 6/1/2023)
  • Numpy version: 1.23.5
  • Opencv version: 4.7.0

Data preprocessing and noise generation

" Before training, you have to rescale the images to 180x180 and adding noise to them. The folder structure is supposed to be:

./data/train/original  for the 180x180 original train images
./data/train/noisy  for the 180x180 noisy train images
./data/test/original  for the 180x180 original test images
./data/test/noisy  for the 180x180 noisy test images

You need the original files for testing just to calculate the PSNR. You can denoise without original files: just put the noisy files also in ./data/test/original " - https://github.com/wbhu/DnCNN-tensorflow

How to Train

./data/train/original  put the 180x180 original train images here
./data/train/noisy  put the 180x180 noisy train images here
  • Put the images at the right places *The number of images must be 10n. (n is a natural number)
  • Excute cmd
  • Move to the directory that the main.py exists using cd command
  • Enter the command below.
python main.py

How to Test

./data/test/original put the 180x180 original test images here *it's totally ok to put the same image here
./data/test/noisy  put the 180x180 noisy test images here 
  • Put the images at the right places
  • Excute cmd
  • Move to the directory that the main.py exists using cd command
  • Enter the command below.
python main.py --phase test

Releases

No releases published

Languages