Repository files navigation

The goal of this project is to upscale and improve the quality of low resolution images.

Since the code is no longer actively maintained, it will be archived on 2025-01-03.

This project contains Keras implementations of different Residual Dense Networks for Single Image Super-Resolution (ISR) as well as scripts to train these networks using content and adversarial loss components.

The implemented networks include:

• The super-scaling Residual Dense Network described inResidual Dense Network for Image Super-Resolution (Zhang et al. 2018)
• The super-scaling Residual in Residual Dense Network described inESRGAN: Enhanced Super-Resolution Generative Adversarial Networks (Wang et al. 2018)
• A multi-output version of the Keras VGG19 network for deep features extraction used in the perceptual loss
• A custom discriminator network based on the one described inPhoto-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (SRGANS, Ledig et al. 2017)

Read the full documentation at:https://idealo.github.io/image-super-resolution/.

Docker scripts andGoogle Colab notebooks are available to carry training and prediction. Also, we provide scripts to facilitate training on the cloud with AWS andnvidia-docker with only a few commands.

ISR is compatible with Python 3.6 and is distributed under the Apache 2.0 license. We welcome any kind of contribution. If you wish to contribute, please see theContribute section.

• Pre-trained networks
• Installation
• Usage
• Additional Information
• Contribute
• Citation
• Maintainers
• License

When training your own model, start with only PSNR loss (50+ epochs, depending on the dataset) and only then introduce GANS and feature loss. This can be controlled by the loss weights argument.

This is just sample, you will need to tune these parameters.

PSNR only:

loss_weights = {  'generator': 1.0,  'feature_extractor': 0.0,  'discriminator': 0.00}

Later:

loss_weights = {  'generator': 0.0,  'feature_extractor': 0.0833,  'discriminator': 0.01}

If you are having trouble loading your own weights or the pre-trained weights (AttributeError: 'str' object has no attribute 'decode'), try:

pip install'h5py==2.10.0' --force-reinstall

Issue

The weights used to produced these images are available directly when creating the model object.

Currently 4 models are available:

• RDN: psnr-large, psnr-small, noise-cancel
• RRDN: gans

Example usage:

model = RRDN(weights='gans')

The network parameters will be automatically chosen.(seeAdditional Information).

RDN model, PSNR driven, choose the optionweights='psnr-large' orweights='psnr-small' when creating a RDN model.

Low resolution image (left), ISR output (center), bicubic scaling (right). Click to zoom.

RRDN model, trained with Adversarial and VGG features losses, choose the optionweights='gans' when creating a RRDN model.

RRDN GANS model (left), bicubic upscaling (right).
->more detailed comparison

RDN model, trained with Adversarial and VGG features losses, choose the optionweights='noise-cancel' when creating a RDN model.

Standard vs GANS model. Click to zoom.

RDN GANS artefact cancelling model (left), RDN standard PSNR driven model (right).
->more detailed comparison

There are two ways to install the Image Super-Resolution package:

• Install ISR from PyPI (recommended):

pip install ISR

• Install ISR from the GitHub source:

git clone https://github.com/idealo/image-super-resolutioncd image-super-resolutionpython setup.py install

Load image and prepare it

importnumpyasnpfromPILimportImageimg=Image.open('data/input/test_images/sample_image.jpg')lr_img=np.array(img)

Load a pre-trained model and run prediction (check the prediction tutorial under notebooks for more details)

fromISR.modelsimportRDNrdn=RDN(weights='psnr-small')sr_img=rdn.predict(lr_img)Image.fromarray(sr_img)

To predict on large images and avoid memory allocation errors, use theby_patch_of_size option for the predict method, for instance

sr_img = model.predict(image, by_patch_of_size=50)

Check the documentation of theImageModel class for further details.

Create the models

fromISR.modelsimportRRDNfromISR.modelsimportDiscriminatorfromISR.modelsimportCut_VGG19lr_train_patch_size=40layers_to_extract= [5,9]scale=2hr_train_patch_size=lr_train_patch_size*scalerrdn=RRDN(arch_params={'C':4,'D':3,'G':64,'G0':64,'T':10,'x':scale},patch_size=lr_train_patch_size)f_ext=Cut_VGG19(patch_size=hr_train_patch_size,layers_to_extract=layers_to_extract)discr=Discriminator(patch_size=hr_train_patch_size,kernel_size=3)

Create a Trainer object using the desired settings and give it the models (f_ext anddiscr are optional)

fromISR.trainimportTrainerloss_weights= {'generator':0.0,'feature_extractor':0.0833,'discriminator':0.01}losses= {'generator':'mae','feature_extractor':'mse','discriminator':'binary_crossentropy'}log_dirs= {'logs':'./logs','weights':'./weights'}learning_rate= {'initial_value':0.0004,'decay_factor':0.5,'decay_frequency':30}flatness= {'min':0.0,'max':0.15,'increase':0.01,'increase_frequency':5}trainer=Trainer(generator=rrdn,discriminator=discr,feature_extractor=f_ext,lr_train_dir='low_res/training/images',hr_train_dir='high_res/training/images',lr_valid_dir='low_res/validation/images',hr_valid_dir='high_res/validation/images',loss_weights=loss_weights,learning_rate=learning_rate,flatness=flatness,dataname='image_dataset',log_dirs=log_dirs,weights_generator=None,weights_discriminator=None,n_validation=40,)

Start training

trainer.train(epochs=80,steps_per_epoch=500,batch_size=16,monitored_metrics={'val_PSNR_Y':'max'})

You can read about how we trained these network weights in our Medium posts:

• part 1:A deep learning based magnifying glass
• part 2:Zoom in... enhance

The weights of the RDN network trained on theDIV2K dataset are available inweights/sample_weights/rdn-C6-D20-G64-G064-x2/PSNR-driven/rdn-C6-D20-G64-G064-x2_PSNR_epoch086.hdf5.
The model was trained usingC=6, D=20, G=64, G0=64 as parameters (see architecture for details) for 86 epochs of 1000 batches of 8 32x32 augmented patches taken from LR images.

The artefact can cancelling weights obtained with a combination of different training sessions using different datasets and perceptual loss with VGG19 and GAN can be found atweights/sample_weights/rdn-C6-D20-G64-G064-x2/ArtefactCancelling/rdn-C6-D20-G64-G064-x2_ArtefactCancelling_epoch219.hdf5We recommend using these weights only when cancelling compression artefacts is a desirable effect.

The main parameters of the architecture structure are:

• D - number of Residual Dense Blocks (RDB)
• C - number of convolutional layers stacked inside a RDB
• G - number of feature maps of each convolutional layers inside the RDBs
• G0 - number of feature maps for convolutions outside of RDBs and of each RBD output

source:Residual Dense Network for Image Super-Resolution

The main parameters of the architecture structure are:

• T - number of Residual in Residual Dense Blocks (RRDB)
• D - number of Residual Dense Blocks (RDB) insider each RRDB
• C - number of convolutional layers stacked inside a RDB
• G - number of feature maps of each convolutional layers inside the RDBs
• G0 - number of feature maps for convolutions outside of RDBs and of each RBD output

source:ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks

We welcome all kinds of contributions, models trained on different datasets, new model architectures and/or hyperparameters combinations that improve the performance of the currently published model.

Will publish the performances of new models in this repository.

See theContribution guide for more details.

To bump up the version, use

bumpversion {part} setup.py

Please cite our work in your publications if it helps your research.

@misc{cardinale2018isr,title={ISR},author={Francesco Cardinale et al.},year={2018},howpublished={\url{https://github.com/idealo/image-super-resolution}},}

• Francesco Cardinale, github:cfrancesco
• Dat Tran, github:datitran

SeeLICENSE for details.