CNDNet: A Cross-State Neural Decoding Network for Unified Semantic Decoding From Perceptual and Imagery EEG

Abstract

Electroencephalography (EEG), owing to its noninvasiveness, high temporal resolution, and the central role of perception and imagery in human cognition, has made EEG-based neural decoding highly significant in fields such as cognitive science and brain–computer interfaces. However, most existing studies focus solely on semantic decoding from either visual perception or visual imagery EEG, without considering their shared neural circuits and semantic invariance. As a result, the two types of EEG have not been placed within a unified framework, greatly limiting the potential of current decoding methods. To address this challenge, a Cross-State Neural Decoding Network (CNDNet) is proposed, where the semantic discrepancy between the two types of EEG signals is minimized as a training objective, enabling the extraction of essential semantic features and realization of cross-state semantic decoding. First, a Feature Extractor based on a spatial–frequency convolutional architecture is employed to extract spatial–spectral features, with a weighted spatial attention mechanism used to enhance semantic components. Then, a Category Classifier combining Group Normalization and Kolmogorov–Arnold Networks is applied to ignore differences in semantic feature representations and capture invariant semantics. Finally, a Domain Discriminator with a similar structure to the Category Classifier is used to determine the domain category of features. CNDNet, validated on two public datasets, achieves an accuracy gain of up to 16.60 percentage points over leading models, demonstrating efficient and robust cross-state neural decoding.