NeuroVision: EEG-to-image reconstruction via progressive neural encoding and cross-modal distillation

Abstract

Reconstructing visual imagery from electroencephalography (EEG) signals represents a fundamental challenge in brain-computer interfaces, as EEG recordings capture rich neural information about visual perception but lack the spatial resolution necessary for direct image reconstruction. Existing approaches typically employ diffusion models to synthesize images from EEG embeddings, yet these methods often fail to preserve the semantic content encoded in neural signals due to information loss during the encoding-decoding process. In this paper, we propose NeuroVision, a novel framework that adapts neural signal encoders for unified EEG understanding and visual reconstruction through progressive neural encoding and cross-modal distillation. Our approach addresses the fundamental trade-off between preserving EEG semantic information and achieving high-quality image reconstruction by introducing a three-stage progressive training scheme that gradually enhances reconstruction capability while maintaining original neural signal understanding. We develop a temporal-spatial attention fusion mechanism to capture multi-channel temporal dynamics in EEG signals, coupled with adaptive feature alignment that dynamically maps EEG representations to visual feature spaces. Furthermore, we introduce a semantic-preserving loss function that ensures reconstructed images faithfully reflect the semantic content of neural activity rather than generating visually plausible but semantically inconsistent outputs. Extensive experiments demonstrate that NeuroVision achieves superior reconstruction quality compared to existing diffusion-based approaches while significantly better preserving semantic correspondence between neural signals and visual content, establishing a new paradigm for EEG-to-image reconstruction that prioritizes semantic fidelity alongside visual quality.