Noise2Noise-for-Cryo-EM-image-denoising

pytorch implementation of noise2noise for Cryo-EM image denoising https://arxiv.org/abs/1803.04189

Network Architecture

Similar to the noise2noise paper image

Loss function

L2 loss

Dependencies

pytorch CUDA 9.0 CuDNN 7.0 Anaconda(python3.6)

Training

python train.py (you need to modify the path in the config.py)

Testing

python test.py (you need to modify the path in the config.py)

Results on the natrual imgs

image train the network using 256x256-pixel crops drawn from the 5k images in the COCO 2017 validation set for 120 epoch. We furthermore randomize the noise standard deviation σ= [0,50] separately for each training example.

Results on Cryo-EM data

image

We train the network using 640*640 crops drawn from the 250 images for 500 epoch for each protein sample dataset. we tested on 2 protein sample dataset,one is aldolase, the other is apoferritin