Taylor expansion-based Kolmogorov-Arnold Network for Blind Image Quality Assessment

🔥 News

• Aug 28, 2025: 🎉 Our new paper is accepted to Journal of Visual Communication and Image Representation (JVCIR), SCI, JCR Q2.
• May 27, 2025: We’ve updated the paper to further explore TaylorKAN and add extensive experiments.
• Dec 21, 2024: 🎉 Our paper is accepted to ICASSP 2025!
• Oct 28, 2024: We release our code.
• Sep 12, 2024: We release our paper on arXiv.

Setup environment

# 1. Clone the repository
git clone https://github.com/CUC-Chen/KAN4IQA.git
cd KAN4IQA

# 2. (Optional) Create and activate Conda environment
conda create -n kan4iqa python=3.10 -y
conda activate kan4iqa

# 3. Install dependencies
pip install -r requirements.txt

Prepare features

Before training and evaluation, you need to prepare the input features. We provide two options:

1. You can choose to manually extract features using ResNet50.
2. We provide pre-extracted features for convenience.
   • Download the five .csv files from this Google Drive link.
   • Place these five .csv files directly into the scripts folder within your cloned repository.

Train and evaluation

You can train TaylorKAN by running:

cd scripts
python train.py

All the results are saved at script/outputs.

Results

Model	BID		CLIVE		KonIQ		SPAQ		FLIVE
	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC
SVR	0.834	0.842	0.754	0.691	0.862	0.850	0.834	0.842	0.537	0.510
MLP	0.750	0.780	0.637	0.554	0.808	0.763	0.855	0.860	0.370	0.319
KAN	0.746	0.756	0.720	0.681	0.772	0.739	0.812	0.810	0.419	0.376
FastKAN	0.813	0.794	0.731	0.699	0.805	0.778	0.845	0.841	0.504	0.413
ChebyKAN	0.751	0.736	0.679	0.549	0.749	0.716	0.800	0.792	0.484	0.344
JacobiKAN	0.762	0.762	0.721	0.628	0.782	0.751	0.811	0.807	0.495	0.407
WavKAN	0.767	0.735	0.752	0.676	0.810	0.777	0.792	0.784	-0.006	0.010
HermiteKAN	0.696	0.656	0.614	0.575	0.737	0.702	0.802	0.800	0.447	0.369
BSRBFKAN	0.793	0.763	0.737	0.692	0.817	0.792	0.846	0.841	0.517	0.430
FourierKAN	0.337	0.358	0.052	0.054	0.096	0.092	0.314	0.274	-0.049	-0.028
EfficientKAN	0.657	0.694	0.731	0.684	0.779	0.752	0.742	0.748	0.473	0.424
TaylorKAN (ours)	0.842	0.843	0.783	0.753	0.830	0.816	0.856	0.862	0.539	0.484

📚 Citation

If you find this work useful for your research, please consider citing our paper:

@inproceedings{yu2025exploring,
  title={Exploring Kolmogorov-Arnold networks for realistic image sharpness assessment},
  author={Yu, Shaode and Chen, Ze and Yang, Zhimu and Gu, Jiacheng and Feng, Bizu and Sun, Qiurui},
  booktitle={ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2025},
  organization={IEEE}
}

@article{chen2025taylor,
  title = {Taylor expansion-based Kolmogorov–Arnold network for blind image quality assessment},
  author = {Chen, Ze and Yu, Shaode},
  journal = {Journal of Visual Communication and Image Representation},
  volume = {112},
  pages = {104571},
  year = {2025},
  issn = {1047-3203},
  doi = {10.1016/j.jvcir.2025.104571},
  url = {https://www.sciencedirect.com/science/article/pii/S1047320325001853}
}