abhijithpunnappurath/a-little-bit-more

Official repository for "A little bit more: Bitplane-wise bit-depth recovery" [IEEE] [arXiv]

BibTeX

@ARTICLE{9606525,
  author={Punnappurath, Abhijith and Brown, Michael S},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={A Little Bit More: Bitplane-Wise Bit-Depth Recovery}, 
  year={2021},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TPAMI.2021.3125692}}

---

Requirements

This code was tested on

• Python 3.6
• Tensorflow 1.12.0
• Keras 2.2.4
• scikit-image 0.15.0
• opencv 3.3.1

Or create a new conda environment with

conda env create -f environment.yml

and activate it with

conda activate bitmore

---

Testing code

Sample usage

• To test 4 to 8-bit recovery on the Kodak dataset (which has already been downloaded to ./data/Test/Kodak) using our D16 model, run

  python test.py --set_names Kodak --type_8_or_16 0 --quant 4 --quant_end 8 --dep 16 
  

  (Note: Set --type_8_or_16 0 for 8-bit images and --type_8_or_16 1 for 16-bit images according to the corresponding --set_names folder)

• To test 6 to 16-bit recovery on this sample image from the Sintel dataset (which has already been downloaded to ./data/Test/Sintel_sample) using our D4 model and save the result, run

  python test.py --set_names Sintel_sample --type_8_or_16 1 --quant 6 --quant_end 16 --dep 4 --save_result 1  
  

• To test 4 to 8-bit recovery on the Kodak dataset and on the sample image from the Sintel dataset using our D4 model, run

  python test.py --set_names Kodak,Sintel_sample --type_8_or_16 0,1 --quant 4 --quant_end 8 --dep 4 
  

---

Main paper results

1. To reproduce the numbers in Table I, run this code to download the data, and this code to produce the outputs.
2. To reproduce the numbers in Table II, run this code to download the data, and this code to produce the outputs.
   • Note: Downloading and preparing Adobe MIT test data can take a while!
3. To reproduce the numbers in Table III, follow these instructions to download the data, and run this code to produce the outputs.
4. To reproduce the numbers in Table IV on the Kodak dataset which has already been downloaded to ./data/Test/Kodak, run this code.
5. To reproduce the numbers in Table V, follow these instructions to download the data, and run this code to produce the outputs.

---

Supplementary material results

1. To reproduce the numbers in Table S4, follow these instructions to download the data, and run this code to produce the outputs.
2. To reproduce the numbers in Table S5, follow these instructions to download the data, and run this code to produce the outputs.
3. To reproduce the numbers in Table S6, and run this code.
4. To reproduce the numbers in Table S7, and run this code.
5. To reproduce the numbers in Table S8, follow these instructions to download the data, and run this code to produce the outputs.

---

Training code

• Run this code to download training data.
  • Note: Downloading and preparing training data can take a while!
• To train a model that predicts the 5^th bit, run

  python train.py --quant 4 
  

  • Note: The images are quantized to 4 bits, and the model is trained to predict the (4+1)^th bit.
• To perform 4 to 8-bit recovery, train four separate models as

  python train.py --quant 4 
  python train.py --quant 5 
  python train.py --quant 6 
  python train.py --quant 7 
  

• The number of residual units is set to 4 by default, i.e., D4 model. To train the D16 model, pass --dep 16 as argument.
• The models are saved to ./models.