Iterative Residual Network for Deep Joint Image Demosaicking and Denoising

The same code can be used for Deep Image Demosaicking using a Cascade of Convolutional Residual Denoising Networks (ECCV 2018)

Filippos Kokkinos [email protected]

Stamatis Lefkimmiatis

Installation and dependencies

This code uses a collection of external packages. You can install the python packages used during development via conda:

conda create --name <env> --file requirements.txt

Processing RAW files

In order to process RAW files take a look at the Jupyter notebook. The algorithm works with RGGB Bayer images, therefore you may need to shift one pixel the images

To produce a comparable output from DCRaw's demosaicking algorithm run:

dcraw -j -T -o 0 +M -w -4 {filename}

Note: We will upload a revised version of Application which will works with all Bayer variants and it will also support Fuji Xtrans.

Models

We provide four pre-trained models in the pretrained_models/ directory. bayer_toy,xtrans,bayer have been trained with no noise, bayer_noise) with approximate noise variances in the range [0, 10] (out of 255).

Training a new model

To train a model for MSR dataset which will be evaluated on all datasets, run:

CUDA_VISIBLE_DEVICES=1 python -B main.py -depth 5 -epochs 100 -gpu -max_iter 10 -k1 5 -k2 5 -save_path experiment1/

We provide dataloader for MSR, MIT, Kodak and McM datasets, therefore you can train and evaluate on any dataset. Furthermore, we also provide training scripts for Fuji XTrans CFA.

usage: main.py [-h] -epochs EPOCHS [-depth DEPTH] [-init] [-save_images]
               [-save_path SAVE_PATH] [-gpu] [-num_gpus NUM_GPUS]
               [-max_iter MAX_ITER] -batch_size BATCH_SIZE [-lr LR] [-k1 K1]
               [-k2 K2] [-clip CLIP] [-estimate_noise]

Joint Demosaick and denoising training script

optional arguments:
  -h, --help            show this help message and exit
  -epochs EPOCHS        Number of epochs
  -depth DEPTH          Depth of ResDNet
  -init                 Initialize input with Bilinear Interpolation
  -save_images
  -save_path SAVE_PATH  Path to save model and results
  -gpu
  -num_gpus NUM_GPUS
  -max_iter MAX_ITER    Total number of iterations to use
  -batch_size BATCH_SIZE
  -lr LR
  -k1 K1                Number of iterations to unroll
  -k2 K2                Number of iterations to backpropagate. Use the same
                        value as k1 for TBPTT
  -clip CLIP            Gradient Clip
  -estimate_noise       Estimate noise std via WMAD estimator

Known issues

• Application currently does not support Fuji XTrans because dcraw rotates for unknown reason the document files
• DCRAW for some RAW images will misinterpret the CFA pattern indicating that it is RGGB, therefore the end result will be wrong. For these images please proceed with manual padding.

Running into issues?

Contact Filippos Kokkinos [email protected]