super-resolution-service's Introduction

Super Resolution

Repository for the super resolution service on the SingularityNET.

This service uses convolutional neural networks to increase the resolution of an image by reconstructing rather than simply resizing it. The images are upscaled by a factor of 4.

This repository was forked from xinntao/ESRGAN. The original code is written in Python 3 (using Pytorch) and has been integrated into the SingularityNET using Python 3.6.

Refer to:

The User's Guide: for information about how to use this code as a SingularityNET service;
The Original Repository: for up-to-date information on xinntao implementation of this code.
SingularityNET Wiki: for information and tutorials on how to use the SingularityNET and its services.

Contributing and Reporting Issues

Please read our guidelines before submitting an issue. If your issue is a bug, please use the bug template pre-populated here. For feature requests and queries you can use this template.

Authors

Ramon Durães - Maintainer - SingularityNET

License

The original repository is licensed under the Apache License v2.0. See the LICENSE file for details.

super-resolution-service's People

Contributors

Stargazers

Watchers

super-resolution-service's Issues

Currently using only proSR x2, x4 and x8 models for the service but proSR_GAN x4 and x8 models have also been provided and, according to the authors, generate images with more details (although performance metrics are slightly worse).

Create UI

Should be fairly straightforward once style-transfer-service's UI is integrated.

Fix issues with temporary files

Make sure images have been properly downloaded;
Make sure all temp files are deleted before exiting (even if an error occurs);
Make sure input and output file names always match.

Explore overfitting the model for the input image

[DONE] Crop input images and create a small database using the crops;
Train model (in a transfer learning fashion) on the crops;
Evaluate the output both visually and using the metrics used in the article to assess results.

Recommend Projects