Decoding visual object recognition from EEG signals

Brain–computer interfaces (BCIs) and clinical EEG require compact and interpretable decoders, yet scalp sensors mix cortical signals and blur frequency-specific activity. Identifying which cortical regions and features carry discriminative visual information enables efficient, anatomically grounded object recognition decoding. This study localizes the cortical sources of informative EEG signals and identifies compact, mechanism-guided features that are most efficient given fixed data or compute budgets. To address this, we construct a source-space decoding pipeline that projects sensor signals onto anatomically defined cortical regions. Trial-wise activity is summarized within regions of interest (ROIs), and four feature families are extracted from each ROI: band-limited power (delta–gamma), line length (LL) for transient activity, temporal morphology, and couplings reflecting coordination between regions. Per-participant Random Forest (RF) classifiers are trained, and generality is quantified as consistency and ROI importance rankings across participants. A low-dimensional representation based on line length yields the strongest overall performance, while temporal morphology and coupling features contribute less under short RSVP (Rapid Serial Visual Presentation) trials. Relative to the EEG-ImageNet sensor-space baseline (310 features), the 24-ROI LL-only stack shows higher reported mean accuracy while using 92% fewer features (24 features), while a finer-grained, extended visual-pathway ROI set shows higher reported mean accuracy while using 84% fewer features (50 features). Adding a small, anatomically constrained high- γ block produces near-tied performance rather than a consistent improvement. These findings indicate that, for single-trial 0.5 s RSVP decoding, most discriminative information is captured by simple time-domain structure in anatomically defined ROIs. High- γ power remains a useful reference feature family, but its incremental value is limited once LL is included. By grounding features in neuro-informed regions, this approach compares favorably, at the level of reported mean accuracy, with the sensor-space baseline while providing clear anatomical attribution at substantially lower dimensionality, supporting lightweight and interpretable EEG decoding.