This is the pytorch implementation of Texture based Super Resolution Networks (TSRN) presented in the paper The Unreasonable Effectiveness of Texture Transfer for Single Image Super-resolution. The code is tested on Pytorch > v0.4 with python 3.5.
If you use this work then cite us.
@inproceedings{waleed2018unreasonable,
title={The unreasonable effectiveness of texture transfer for single image super-resolution},
author={Waleed Gondal, Muhammad and Scholkopf, Bernhard and Hirsch, Michael},
booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
pages={0--0},
year={2018}
}
Install pytorch and torchvision first from the official website. You can use the following command to install the latest version for python 3.5.
pip3 install torch torchvision
For training the semantically guided TSRN models, we used MS-COCO Stuff dataset. Semantically annotated dataset can be downloaded from this website.
Two different training configurations are provided for TSRN-global and TSRN-segmented. To train a TSRN-Global model provide path to root directory and set upscale factor accordingly
python train.py --upscale_factor 8 --cropsize 256 --path_data '/home/mscoco/train2017'
To train a TSRN-Segmented model provide additional path to annotation files for semantic segmentation.
python train_segmented.py --upscale_factor 4 --cropsize 256 --path_data '/home/mscoco/train2017' --annFile '/home/coco_stuff/stuff_train2017.json'
The weights and sample super-resolved images are saved at each epoch in the trainings folder.
Pretrained models of TSRN-S and TSRN-G for both 4x and 8x SISR are provided in the resources folder. The models should be able to run on both CPU and GPU. To test a model, provide path to the test images, the upscaling factor and the corresponding path to the pretrained weights.
python test.py --upscale_factor 4 --path_model './resources/pretrained/tsrn_global_4x.pth' --path_data './resources/images'