A bilinear feature extraction method for rapid serial visual presentation triage

Searching for target objects in large volume imagery is a challenging problem, and the rapid serial visual presentation (RSVP) triage based on the detection of event-related potentials (ERP) is potentially a promising solution to the problem. Due to the fact that ERP elicited by targets and those by non-targets differ not only on spatial patterns but temporal patterns, this paper proposes a feature extraction method namely bilinear common spatial pattern (BCSP), which is designed to capture discriminative spatio-temporal features of ERP for triage classification. The proposed method extends common spatial pattern (CSP) by incorporating the core idea of bilinear discriminant analysis (BDA) into it. Specifically, in addition to the spatial filters which CSP also looks for, BCSP acquires the temporal filters for learning the temporal patterns of ERP. Both the spatial filters and temporal filters are obtained by using the same Ramoser´s technique, but in an iterative manner. With discriminative temporal information involved, BCSP has manifested remarkable advantages in RSVP triage experiments, as demonstrated by a significant increase of 11.2% in average classification accuracy in comparison with CSP, with p <; 0.001 in paired t-test.