Unsupervised Image Deraining: Optimization Model Driven Deep CNN Paper

By Changfeng Yu*, Yi Chang* (https://scholar.google.com/citations?user=I1nZ67YAAAAJ&hl=en)(* indicates equal contribution)

Demo

Dependencies and Installation

• Python 3 (Recommend to use Anaconda)
• PyTorch >= 1.2
• NVIDIA GPU + CUDA
• Python packages: pip install -r enviroment.txt

Dataset

• Our dataset RealRain(including the angle) and Rain_cityscape can be downloaded here.
• Make your own data:
  1. Simulating by yourself, the angle information can be easily obtained during the simulation.
  2. Other sythetic datasets or real image dataset, you need to label the angle informations by ./lib/angle_label.py
• the dataset should have the following structure:

  -train
  -rain/data
  -angle/data
  -clean/data
  

  -test
  -rain/data
  -angle/data
  

• the training data based cityscape can be downlowed form Baidu Netdisk the extraction code is hust

How to Train

• UDGNet
  1. Run command:

  python train_Decomposition_angle.py --rain_path ./dataset/test/rain --angle_path ./data/test/angle --clean_path ./data/test/rain --reset 1
  

How to Test

• UDGNet
  1. Run command:

  python Test_Decomposition_angle.py --rain_path ./dataset/test/rain --angle_path ./data/test/angle --clean_path ./data/test/rain --weight_path ./output/real_model/generator_backup.pth
  

UDGNet