FSDBN achieves 69% top-1 accuracy in zero-shot brain-to-image retrieval

A team of researchers from Nanjing University of Aeronautics and Astronautics has unveiled FSDBN, a novel framework for EEG-based visual decoding that significantly improves the ability to reconstruct or retrieve images directly from brain activity. The key innovation is foreground-aware alignment: existing methods treat all visual content equally, but the human brain prioritizes salient foreground objects over background clutter. FSDBN introduces Semantic-Consistent Saliency Alignment to isolate semantically relevant regions, then uses a Semantic-Prior Dynamic Gating Fusion module to adaptively weigh foreground and background contributions. On the EEG side, the system models signals as dynamic spatiotemporal brain networks whose functional connectivity reorganizes in real time to reflect attention to salient features.

In zero-shot brain-to-image retrieval experiments—where the model must match EEG recordings to images it has never seen during training—FSDBN achieves 69.0% top-1 accuracy and 92.2% top-5 accuracy, both state-of-the-art. This represents a major leap in non-invasive brain-computer interfaces, with potential applications in rapid image search, assistive communication, and neuroscience research. The code is publicly available on GitHub, enabling further development and replication. The work highlights how combining visual saliency priors with dynamic brain network modeling can overcome the perceptual asymmetry that has long plagued EEG-based visual decoding.