On the Role of Low-Level Visual Features in EEG-Based Image Reconstruction

Decoding visual stimuli from electroencephalography (EEG) has gained increasing attention in recent years due to EEG's affordability, portability, and high temporal resolution. State-of-the-art approaches typically map EEG signals to Contrastive Language Image Pre-training (CLIP) embeddings via contrastive learning, followed by image reconstruction using conditional diffusion models. These methods were originally developed for other imaging techniques and mainly focus on reconstructing high-level semantic information, often neglecting low-level visual features such as color, texture, and posture, especially when applied to EEG signals. In this study, we explore a two-level decoding framework designed to incorporate low-level visual information alongside highlevel representations. Compared to the Adaptive Thinking Mapper (ATM) baseline, the current state-of-the-art method for EEG decoding, our model demonstrates improved recovery of coarse information from EEG but continues to struggle with fine-grained details and semantic completeness. Furthermore, integrating the low-level and high-level signals into a stable diffusion technique revealed improved preservation of the low-level details and degradation of semantic coherence. These findings underscore the need for more effective low-level modeling and suggest that further exploration of the two-level decoding framework is required to better leverage EEG signals for image reconstruction. Code: https://github.com/jyang635/EEG_decoding.git