Zero-shot visual decoding and image retrieval method based on non-invasive electroencephal signal
By combining multivariate noise normalization whitening and fixed-wavelength removal with a spatiotemporally coupled block attention network, the problems of noise interference and common evoked components in EEG signals are solved, achieving high-precision zero-sample visual decoding and image retrieval, and improving the decoding performance and open-world generalization ability of the brain-computer interface system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGNAN UNIV
- Filing Date
- 2026-03-12
- Publication Date
- 2026-07-17
AI Technical Summary
Existing non-invasive EEG visual decoding methods struggle to effectively extract high-order specific semantic features related to visual content from strong noise and common evoked potential interference, resulting in a large cross-modal semantic gap and poor zero-sample generalization ability.
Spatial covariance noise and common evoked components of EEG signals are eliminated through multivariate noise normalization whitening and fixed-wavelength removal. Features are extracted by combining spatiotemporally coupled block attention networks, and cross-modal alignment is performed using a pre-trained visual-language large model. A contrastive loss based on InfoNCE is constructed for training.
It significantly improves the accuracy and robustness of zero-shot visual decoding and image retrieval, achieving high-precision open-world image retrieval capabilities and bridging the modal gap between EEG signals and visual images.
Smart Images
Figure CN122412973A_ABST