zogenzeng / koniq

KonIQ-10k Deep Learning Models

NR-IQA models trained on the KonIQ-10k dataset

This is part of the code for the paper "KonIQ-10k: An ecologically valid database for deep learning of blind image quality assessment". The included Python 2.7 notebooks rely on the kutils library. The Google colab requires the ku library. Project data is available for download from osf.io.

To quickly try out the Koncept512 model:

pip install koncept

Please cite the following paper if you use the code or package:

@article{koniq10k,
author={V. {Hosu} and H. {Lin} and T. {Sziranyi} and D. {Saupe}},
journal={IEEE Transactions on Image Processing},
title={KonIQ-10k: An Ecologically Valid Database for Deep Learning of Blind Image Quality Assessment},
year={2020},
volume={29},
pages={4041-4056}}

Overview

Google colab notebook, Python 3 compatible:

koncept512_train_test_py3_with_kuti.ipynb (updated Sept 2021)

• Download KonIQ-10k dataset, train the KonCept512 model and test it
• Load a pre-trained KonCept512 model, and use it to predict the quality of an image

Python 2.7 notebooks:

train_koncept512.ipynb, train_koncept224.ipynb:

• Training and testing code for the KonCept512 and KonCept224 model (on KonIQ-10k).
• Ready-trained model weights for KonCept512 and KonCept224.

train_deeprn.ipynb

• Reimplementation of the DeepRN model trained on KonIQ-10k, following the advice of the original author, Domonkos Varga.
• Re-trained model weights (on SPP features) are available here.
• The features extracted from KonIQ-10k are available here.

metadata/koniq10k_distributions_sets.csv

• Contains image file names, scores, and train/validation/test split assignment (random).