A CNN-Based Blind Denoising Method

Official implementation of the BioCAS 2019 paper: A CNN-Based Blind Denoising Method for Endoscopic Images

Pytorch implementation of blind denoising network and keras implementation of blind image quality assessment network.

For blind denoising, we optimize the Deep Image Prior method by transfer learning to reduce the number of iteration.

To determine the quality of reconstructed image, the blind image assessment network based on MobileNet is presented to estimate the scores of image quality.

Architecture

Dependencies

• torch (1.0.1.post2)
• tensorflow (1.12.0)
• keras (2.1.6)
• numpy
• matplotlib
• skimage
• opencv-python
• pandas

Run on ubuntu 16.04, python3.5, CUDA 9.0, cuDNN 7.5 with 8 Nvidia TITAN Xp GPUs.

Usage

Train blind denoising network(BDN) to reconstruct noisy image

Step1. Reconstruct PolyU dataset

python run_BDN.py --dataset_real_path dataset/CC/real/ \

​ --dataset_mean_path dataset/CC/mean/ \

​ --output_path dataset/CC/20200111 \

​ --num_iter 3500 --save_iter 20 --lr 1e-2 --gpu_id 1

Step2. Reconstruct CC dataset

python run_BDN.py --dataset_real_path dataset/PolyU/real/ \

​ --dataset_mean_path dataset/PolyU/mean/ \

​ --output_path dataset/PolyU/20200111 \

​ --num_iter 5000 --save_iter 20 --lr 1e-2 --gpu_id 2

Tips

PolyU dataset has 100 pairs of noisy and clean images and CC dataset has 15 pairs of noisy and clean images.

Use the method of deep image prior to reconstruct one image will cost about 15 minutes with 3000 iteration.

So you can simply download the reconstructed images from here or here and then update the dataset.

Train blind image quality assessment network(BIQAN)

Setp1. Convert the PSNR of reconstructed images to AVA dataset format

python generator_dataset.py --dataset_path dataset/PolyU_Mulit_UN_GN \

​ --output_path dataset/PolyU_Mulit_UN_GN_PNSR.txt

Step2. Train the blind image quality assessment network

python train_BIQAN.py --dataset_image_path dataset/PolyU_Mulit_UN_GN \

​ --dataset_file_path dataset/PolyU_Mulit_UN_GN_PNSR.txt \

​ --output_checkpoint checkpoint/PolyU_Mulit_UN_GN/mobilenet \

​ --output_path result/PolyU_Mulit_UN_GN \

​ --epochs 40 --lr 1e-3 --gpu_id 3

Tips

You can download the trained model from here or here

Step3. Test the blind image quality assessment network

python test_BIQAN.py --dataset_image_path dataset/CC_Resume_All32 \

​ --pre_trained_model checkpoint/PolyU_Mulit_UN_GN/mobilenet/weights.004-0.046.hdf5 \

​ --output_path result/CC_Resume_All32 \

​ --start_index 90 --gpu_id 4

Results

Some reconstructed images

Mobilenet training loss

Denoised effect

clean image	noisy image	denoised image

Comparison of CNN based algorithms

Method	Input	Label	Output	Back-Bone
MLP	Clean image + Gaussian noise	Clean image	Denoised image	Fully connected network
DnCNN	Clean image + Gaussian noise	Gaussian noise	Estimated noise	[Conv + BN + RELU]
Noise2Noise	Clean image + Gaussian noise_A	Clean image + Gaussian noise_B	Denoised image	Encoder-Decoder/U-net
Deep image Prior	Gaussian noise + Uniform noise	Noisy Image	Reconstructed image	Encoder-Decoder/U-net

Quantitative results using PSNR on the CC dataset

Camera	CBM3D	NC	NI	CC	Proposed
Canon 5D ISO = 3200	37.79	37.72	38.75	38.37	40.83
	34.34	35.26	35.57	35.37	36.65
	34.27	34.89	35.55	34.91	36.30
Nikon D600 ISO = 3200	33.70	34.70	35.59	34.98	34.97
	34.33	34.32	36.78	35.95	35.88
	35.75	38.57	39.30	40.51	41.19
Nikon D800 ISO= 1600	36.15	38.18	38.02	37.99	38.39
	36.57	38.84	38.99	40.36	41.94
	35.47	38.44	38.19	38.30	37.99
Nikon D800 ISO = 3200	34.00	38.22	38.05	39.01	39.76
	33.43	35.72	35.71	36.75	35.59
	33.53	38.58	32.91	39.06	41.22
Nikon D800 ISO = 6400	29.97	33.61	33.51	34.61	34.94
	30.33	32.57	32.75	33.21	33.38
	30.21	32.86	32.88	33.22	33.02
Average	33.99	36.17	35.33	36.88	37.47

Reference

Part of code refers from deep image prior and neural-image-assessment.

Citation

If you find this project useful, we would be grateful if you cite the paper:

@article{BlindDenoising2019,
author = {Shaofeng Zou, Mingzhu Long, Xuyang Wang, Xiang Xie, Guolin Li, Zhihua Wang},
conference = {BioCAS 2019},
title = {A CNN-Based Blind Denoising Method for Endoscopic Images},
url = {https://ieeexplore.ieee.org/abstract/document/8918994/},
year = {2019}
}

License

• For academic and non-commercial use only.
• For commercial use, please contact zousf19@mails.tsinghua.edu.cn