CNN resolution degrader

Convolutional neural networks for downscaling aerial images. Trained on a set of real-life high and low-resolution images from the Proba-V satellite mission. Created as a part of data augmentation and generation for super-resolution project.

This is part of a research project published at:

M. Ziaja, J. Nalepa and M. Kawulok, "Data Augmentation for Multi-Image Super-Resolution," IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium, Kuala Lumpur, Malaysia, 2022, pp. 119-122, doi: 10.1109/IGARSS46834.2022.9884609.

Running the project

The data is managed mostly by the DVC system. To obtain access to the DVC remotes please contact the maintainer: maciejzjg@gmail.com.

There are two ways of managing data, one can either use the full Proba-V dataset for degradation training and then create artificial Sentinel-2 datasets for super-resolution, or split Proba-V in two and use one half for degradation fitting and the other for super-resolution. These splits of Proba-V are denoted as a and b consecutively.

Input datasets

The input dataset generation for training degradation networks is stored in the data_generation_degradation (may require some readjustments to regenerate the datasets). This part of data preprocessing is not managed by DVC pipelines, although the datasets are stored with DVC. They can be downloaded from remote with the dvc pull command.

Networks training

The models are automatically managed depending on changes in source code and configuration in params.yaml. The training pipeline is defined in dvc.yaml. Experiment reproduction is done with the dvc repro command. Pre-trained models created with the current state of the master branch can be pulled from the DVC remote. There are three model architectures for degradation: a simple fully convolutional network, a U-Net, and a GAN network.

Evaluation

Evaluation is handled outside of the DVC pipelining. To test run the python -m cnn_res_degrader.test (use help to get familiar with the usage). The results are presented as heatmap plots. The GAN network doesn't feature early stopping and checkpointing best model, so every epoch the generator network is saved. To examine which epoch performed best on the validation subset use the find_best_gan.py script.

Exporting/data generation

Artificial datasets for super-resolution training can be generated using scripts from the data_generation_superres. Consult the help for usage.

Misc

Some insights and data exploration can be found in Jupyter notebooks in the analysis directory.

The results of super-resolution trainings using artificial datasets and a wider context can be found in a paper in this repository (however, it may be outdated). The work is continued at KP Labs.