DeepGuidedFilteringNetwork

DGF for Image Processing

Prepare Dataset

1. Download MIT-Adobe FiveK Dataset.

   dataset
   ├── ......
   └── fivek
       ├── ......
       └── raw_photos
           ├── HQa1to700
           │   └── photos
           │       ├── a0001-jmac_DSC1459.dng
           │       └── ......
           └── ......

2. Convert from *.dng to *.tif.

   cd scripts
   python convert_dng_to_tif.py

3. Generate Training/Test set with different resolution.

   cd scripts
   
   python precompute_size.py --min 512
   python precompute_size.py --min 1024
   python precompute_size.py --random
   
   python resize_image.py --file_name 512
   python resize_image.py --file_name 1024
   python resize_image.py --file_name random

4. Generate Ground Truth for each task.
   • L₀ smoothing / Multi-scale detail manipulation / Style transfer / Non-local dehazing

     cd scripts/[l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing]
     
     matlab -r "prepare_dataset(   '512'); exit" -nodisplay
     matlab -r "prepare_dataset(  '1024'); exit" -nodisplay
     matlab -r "prepare_dataset('random'); exit" -nodisplay

   • Image retouching (Auto-PS, Auto Beautification)
     1. Generate / Download Ground Truth

        dataset
        ├── ......
        └── fivek
            ├── ......
            └── gts
                ├── a0001-jmac_DSC1459.tif
                └── ......

        • Option 1: Generate ground truth with Adobe Lightroom by following the instruction.
        • Option 2 (Preferred): Download from the website.

          cd scripts/auto_ps
          
          # Expert [A|B|C|D|E]
          bash generate_list.sh [a|b|c|d|e]
          
          cd ../../dataset/fivek && mkdir gts && cd gts
          # Option 1: single thread
          wget -i ../../../scripts/auto_ps/img_list_[a|b|c|d|e].csv
          # Option 2: multi thread
          sudo apt-get install aria2
          aria2c -j 100 -i ../../../scripts/auto_ps/img_list_[a|b|c|d|e].csv

     2. Postprocess

        cd scripts/auto_ps
        python postprocess.py
        
        cd ..
        python resize_image.py --file_name 512    --task auto_ps
        python resize_image.py --file_name 1024   --task auto_ps
        python resize_image.py --file_name random --task auto_ps

5. Split Training/Test set

   cd scripts/training_test_split
   python split.py --set 512 \
                   --task [l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing|auto_ps]
   python split.py --set 1024 \
                   --task [l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing|auto_ps]
   python split.py --set random \
                   --task [l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing|auto_ps]

Training

• Option 1: Train from scratch

  python train_hr.py --task [l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing|auto_ps] \
                     --name [HR|HR_AD|CONV_HR|CONV_HR_AD] \
                     --model [deep_guided_filter|deep_guided_filter_advanced|deep_conv_guided_filter|deep_conv_guided_filter_adv]

• Option 2: Train with low-resolution data + Finetune

  python train_lr.py --task [l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing|auto_ps]
  
  python train_hr_finetune.py --task [l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing|auto_ps] \
                              --name [HR_FT|HR_AD_FT|CONV_HR_FT|CONV_HR_AD_FT] \
                              --model [deep_guided_filter|deep_guided_filter_advanced|deep_conv_guided_filter|deep_conv_guided_filter_adv]

NOTE:

• deep_guided_filter: DGF_b
• deep_guided_filter_advanced: DGF
• deep_conv_guided_filter: DGF_b^c
• deep_conv_guided_filter_adv: DGF^c

Evaluate

python test_hr.py --task [l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing|auto_ps] \
                  --name [LR|HR|HR_AD|CONV_HR|CONV_HR_AD|HR_FT|HR_AD_FT|CONV_HR_FT|CONV_HR_AD_FT] \
                  --model [guided_filter|deep_guided_filter|deep_guided_filter_advanced|deep_conv_guided_filter|deep_conv_guided_filter_adv]

NOTE:

• guided_filter: DGF_s
• deep_guided_filter: DGF_b
• deep_guided_filter_advanced: DGF
• deep_conv_guided_filter: DGF_b^c
• deep_conv_guided_filter_adv: DGF^c

Running Time

python test_time.py --model_id [0|1|2|3|4|5]

NOTE:

• 0: DGF_b
• 1: GuidedFilteringLayer
• 2: DGF
• 3: ConvGuidedFilteringLayer
• 4: DGF_b^c
• 5: DGF^c

Predict

python predict.py  --task [l0_smooth|multi_scale_detail_manipulation|style_transfer|non_local_dehazing|auto_ps] \
                  [--img_path IMG_PATH | --img_list IMG_LIST] \
                   --save_folder SAVE_FOLDER \
                   --model [guided_filter|deep_guided_filter|deep_guided_filter_advanced|deep_conv_guided_filter|deep_conv_guided_filter_adv] \
                   --low_size 64 \
                   --gpu 0 \
                  [--gray]

NOTE:

• --model
  • guided_filter: DGF_s
  • deep_guided_filter: DGF_b
  • deep_guided_filter_advanced: DGF
  • deep_conv_guided_filter: DGF_b^c
  • deep_conv_guided_filter_adv: DGF^c
• --gray: It's better to generate gray images for style transfer.

Acknowledgement

A part of the code was adapted from FastImageProcessing.