royaljava / dsrn

Implementation for our CVPR 2018 paper

Image Super-resolution via Dual-state Recurrent Neural Networks (CVPR 2018)

[Paper Link]

Citation

@inproceedings{han2018image,  
	title={Image super-resolution via dual-state recurrent networks},
	author={Han, Wei and Chang, Shiyu and Liu, Ding and Yu, Mo and Witbrock, Michael and Huang, Thomas S},
	booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
	year={2018}
}

Dependencies

• Common python dependencies can be installed via pip install -r requirements.txt
• Lingvo (for inference only), see linvgo project page for installation instructions.

Data

There is a very helpful repo collected download links for all the training and test sets needed here.

Training

The training data is specified by a file list of HR images. No futher pre-processing is needed as we perform downsampling and augmentation on-the-fly.

Use train.py and the model specification file model_recurrent_s2_u128_avg_t7.py to start a training job.

Inference

We release our models in tensorflow lingvo format such that the models are self contained for inference tasks. Each model consists of by a inference_graph.pbtxt and a checkpoint file.

To run the inference with provided pre-trained models on an image, use provided predictor.py: Example:

`python predictor.py --checkpoint=models/x3/ckpt-00754300 --inference_graph=models/x3/inference.pbtxt --image_path=./cat.png --output_dir=./`

The script will write super-resolved images to output_dir.

Evaluation

Use evaluate.py to compute average PSNR on a test set after saving all the model predicted images. Eval set is also specified by a file list. Example:

`python evaluate.py --hr_flist=flists/set5.list --prediction_dir=${your_pred_dir}`

Acknowledgement

This code is partly based on a previous work from our group [here]