Visual_EEG_Decoding

Leveraging Visual Blur Perception Characteristics for EEG Decoding [AAAI 2026]

Introduction

This is the official implementation for Leveraging Visual Blur Perception Characteristics for EEG Decoding [AAAI 2026].

In this paper, we propose a novel visual decoding framework inspired by human perceptual blurring, achieving a top-1 accuracy of 80% and a top-5 accuracy of 96.9%, surpassing previous state-of-the-art methods by margins of 29.1% and 17.2%, respectively. These findings highlight the potential of incorporating perceptual properties into EEG-based visual decoding.

How to Use

1. Data Preparation

Download the Things-image from the OSF repository, Things-EEG from the OSF repository. We also provide the processed Things-EEG data and the pretrained CLIP model weights on Quark Netdisk.

If you download the processed data directly, you can skip the following processing steps.

Arrange the raw data as follows:

data
└── things_eeg
    ├── Image_set
    │   ├── train_images
    │   └── test_images
    └── Raw_eeg
        ├── sub-01
        ├── ...
        └── sub-10

2. EEG Data Processing

# Set subject number 'sub' from 1 to 10 to process each subject's EEG data.
python preprocess/process_eeg.py --subject sub

3. Image Data Processing

# Generate multi-blur CLIP features.
python preprocess/process_image.py

After the above steps, the directory structure will be:

data
└── things_eeg
    ├── Image_set
    │   ├── train_images
    │   └── test_images
    ├── Image_feature
    │   ├── MultiBlur_RN50_train.pt
    │   ├── MultiBlur_RN50_test.pt
    │   ├── MultiBlur_ViT-H-14_train.pt
    │   └── MultiBlur_ViT-H-14_test.pt
    └── Preprocessed_data
        ├── sub-01
        ├── ...
        └── sub-10

4. Run Training

Intra-subject:

python main_eeg.py

Inter-subject:

python main_eeg.py --cross_subject True

Experimental Results on Things-EEG

Note: The results reported in our AAAI 2026 paper were obtained using CLIP RN50 as the vision encoder. Here we supplement additional experimental results using CLIP ViT-H-14 on the Things-EEG dataset under the same conditions for direct comparison. We report two types of metrics: results selected by the validation set (reflecting practical performance) and the best results on the test set (reflecting the upper bound).

We provide zero-shot retrieval results using MultiBlur ViT-H-14 as the vision encoder under two settings:

• Intra-subject: train and test on the same subject.
• Inter-subject: leave one subject out for test.

The processed MultiBlur ViT-H-14 image features have been added to the Quark Netdisk for download.

For fair comparison: When comparing with our method, please ensure the same model selection strategy is used.

Selected by validation set:

Intra-subject:

Metric	sub-01	sub-02	sub-03	sub-04	sub-05	sub-06	sub-07	sub-08	sub-09	sub-10	Avg
Top-1 Acc	0.655	0.640	0.640	0.705	0.585	0.680	0.555	0.710	0.640	0.710	0.652
Top-3 Acc	0.880	0.855	0.875	0.910	0.805	0.865	0.805	0.900	0.865	0.930	0.869
Top-5 Acc	0.925	0.895	0.915	0.950	0.870	0.905	0.880	0.950	0.930	0.960	0.918

Inter-subject:

Metric	sub-01	sub-02	sub-03	sub-04	sub-05	sub-06	sub-07	sub-08	sub-09	sub-10	Avg
Top-1 Acc	0.210	0.250	0.150	0.160	0.165	0.125	0.180	0.170	0.130	0.290	0.183
Top-3 Acc	0.385	0.430	0.265	0.325	0.320	0.295	0.305	0.290	0.275	0.520	0.341
Top-5 Acc	0.505	0.535	0.350	0.410	0.425	0.385	0.380	0.405	0.355	0.630	0.438

Best metrics (historical best across epochs):

Intra-subject:

Metric	sub-01	sub-02	sub-03	sub-04	sub-05	sub-06	sub-07	sub-08	sub-09	sub-10	Avg
Top-1 Acc	0.685	0.655	0.695	0.725	0.625	0.705	0.620	0.765	0.700	0.765	0.694
Top-3 Acc	0.865	0.830	0.850	0.870	0.790	0.860	0.840	0.925	0.870	0.925	0.863
Top-5 Acc	0.910	0.895	0.910	0.930	0.855	0.910	0.885	0.940	0.915	0.965	0.911

Inter-subject:

Metric	sub-01	sub-02	sub-03	sub-04	sub-05	sub-06	sub-07	sub-08	sub-09	sub-10	Avg
Top-1 Acc	0.255	0.325	0.180	0.205	0.190	0.150	0.200	0.190	0.205	0.325	0.223
Top-3 Acc	0.420	0.465	0.290	0.410	0.355	0.265	0.320	0.320	0.345	0.530	0.372
Top-5 Acc	0.495	0.570	0.350	0.510	0.470	0.365	0.455	0.425	0.430	0.665	0.473

Citation

@inproceedings{liu2026leveraging,
  title={Leveraging Visual Blur Perception Characteristics for EEG Decoding},
  author={Liu, Wenchao and Li, Hongwei and Xu, Zhouyang and Ma, Lin and Li, Haifeng},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={21},
  pages={17580--17588},
  year={2026}
}

If you find this work useful, please consider citing our paper and giving this repository a ⭐!

Acknowledgement

We extend our gratitude to the prior works UBP and NICE-EEG for their pioneering contributions to this field.