Clustering-Based Query Expansion for EEG-Driven Image Retrieval

Decoding visual perception from electroencephalography (EEG) signals through image retrieval is a challenging task due to the low signal-to-noise ratio and limited spatial resolution of scalp-recorded brain activity. While recent methods have achieved promising results in EEG-to-image retrieval, the semantic coherence of the retrieved results remains suboptimal, and the potential for refining EEG embeddings using the structure of the retrieval space has not been fully exploited. This work proposes a clustering-based query expansion method that improves EEG-driven image retrieval without requiring additional model training. Given an initial EEG embedding as a query, the top 20 retrieved candidates are grouped using hierarchical clustering with average linkage, and the cluster containing the top 1 candidate is selected. The embeddings within this cluster are averaged with the original query to produce a refined embedding, which is then used for a second retrieval. The method is evaluated on the THINGS-EEG dataset using two complementary metrics: the CLIP similarity between the retrieved images and the query stimulus, which measures the semantic similarity of the retrieved top-K, and Mean Average Precision at rank 20 with automatic relevance labels. Results show that the method achieves marked improvements in the best cases (CLIP@5 gain up to +0.107, with a mean of +0.066 over the twenty best queries, and MAP@20 gain up to +0.433, with a mean of +0.217), while the worst cases degrade when the initial retrieval is poor. The proposed method offers a computationally lightweight post-processing refinement for EEG-based retrieval systems and provides insights into the role of the retrieval-space structure in improving semantic retrieval.