Abstract : Image restoration is a challenging and ill-posed problem which also has been a long-standing issue. However, most of learning based restoration methods are proposed to target one degradation type which means they are lack of generalization. In this paper, we proposed a multi-branch restoration model inspired from the Human Visual System (i.e., Retinal Ganglion Cells) which can achieve multiple restoration tasks in a general framework. The experiments show that the proposed multi-branch architecture, called CMFNet, has competitive performance results on four datasets, including image deblurring, dehazing and deraindrop which are very common applications for autonomous cars.

Overall Framework of CMFNet
Branch U-Net architecture	Mixed Skip Connection (MSC)

You can simply demo on my space of Hugging Face

• Dehazing
• Deraindrop
• Deblurring

Or test on local environment:
To test the pre-trained models of Deraindrop, Dehaze, Deblurring on your own images, run

python demo.py --input_dir images_folder_path --result_dir save_images_here --weights path_to_models

Here is an example to perform Deraindrop:

python demo.py --input_dir './demo_samples/deraindrop' --result_dir './demo_results' --weights './pretrained_model/deraindrop_model.pth'

All pre-trained models can be downloaded at pretrained_model/README.md or here

To train the restoration models of Deraindrop, Dehaze and Deblurring. You should check the following components:

• training.yaml:

  # Training configuration
  GPU: [0,1,2,3]
  
  VERBOSE: False
  
  MODEL:
    MODE: 'Deblur'
  
  # Optimization arguments.
  OPTIM:
    BATCH: 2
    EPOCHS: 150
    # NEPOCH_DECAY: [10]
    LR_INITIAL: 2e-4
    LR_MIN: 1e-6
    # BETA1: 0.9
  
  TRAINING:
    VAL_AFTER_EVERY: 1
    RESUME: False
    TRAIN_PS: 256
    VAL_PS: 256
    TRAIN_DIR: './datasets/deraindrop/train'       # path to training data
    VAL_DIR: './datasets/deraindrop/test' # path to validation data
    SAVE_DIR: './checkpoints'           # path to save models and images
  
  

• Details of Hyperparameters

   -------------------------------------------------
   GoPro dataset:
   Training patches: 33648 (2103 x 16)
   Validation: 1111
   Initial learning rate: 2e-4
   Final learning rate: 1e-6
   Training epochs: 150 (120 is enough)
  Training time (on single 2080ti): about 10 days
  
   Raindrop dataset:
   Training patches: 6888 (861 x 8)
   Validation: 1228 (307 x 4)
   Initial learning rate: 2e-4
   Final learning rate: 1e-6
   Training epochs: 150 (100 is enough)
  Training time (on single 1080ti): about 2.5 days
  
   IO-Haze dataset:
   Training patches: 15000 (75 x 200)
   Validation: 55
   Initial learning rate: 1e-4
   Final learning rate: 1e-6
   Training epochs: 150 (50 is enough)
  Training time (on single 1080ti): about 3 days
   -------------------------------------------------
  

• Dataset:
  The preparation of dataset in more detail, see datasets/README.md.

• Train:
  If the above path and data are all correctly setting, just simply run:

  python train.py
  

To test the models of Deraindrop, Dehaze, Deblurring with ground truth, see the test.py and run

python test.py --input_dir images_folder_path --result_dir save_images_here --weights path_to_models --dataset type_of_task --gpus CUDA_VISIBLE_DEVICES

Here is an example to perform Deraindrop:

python test.py --input_dir './datasets/' --result_dir './test_results/' --weights './pretrained_model/deraindrop_model.pth' --dataset deraindrop --gpus '0'

To test the PSNR and SSIM of Deraindrop, see the evaluation_Y.py and run

python evaluation_Y.py --input_dir path_to_restored_images --gt_dir path_to_gt_images

Here is an example:

python valuation_Y.py --input_dir './test_results/deraindrop' --gt_dir './demo_samples/deraindrop'

And to test the PSNR and SSIM of Dehaze and Deblur, see the evaluation_RGB.m

Result Tables (Click to expand)

Restoration task	Result Tables
Deraindrop
Dehaze
Deblur

Visual Comparison Figures (Click to expand)

Restoration task	Restored images	Ground Truth
Deraindrop
Dehaze
Deblur

More visual results can be downloaded at here.

If you have any question, feel free to contact [email protected]