Testing the effects of presentation modality and presence of similarity percentages in the Automatic Image Recognition System in Italy

Automatic Face Recognition Systems (AFRS) are commonly used by police operators to compare the image of an unknown subject (target face) with those already recorded in specific databases. In operational forensic workflows, AFRS support identification by generating a lineup through a one-to-many (1:N) search, comparing a probe image against a large biometric gallery in which candidates are ranked by similarity scores. The purpose of the present study was to determine how different interface configurations, specifically image presentation modality and the inclusion of similarity percentages, impact human decision-making and identification accuracy. No study has investigated the impact of different layouts on the accuracy of detecting the correct potential candidate. Here, we tested the effects of presentation modality and the presence/absence of similarity percentages on the recognition performance of untrained participants in an AFRS. Participants studied a target face for 5 seconds and were then asked to recognize it from a list of 50 faces, which may or may not have included the target. The faces were presented in one of three formats: simultaneously (8 images per slide), semi-sequentially (4 images per slide), or sequentially (1 image per slide) and participants were always given the opportunity to compare the selected image with the target face (direct comparison). Each image was either shown with or without a system-generated similarity percentage. When similarity percentages were shown, participants were instructed either to use or ignore them. Results showed that, for target-absent lists, false alarms were fewer with the simultaneous layout than with the semi-sequential or sequential ones. Direct comparisons were also significantly lower with the simultaneous condition, and, in this condition, participants who made more direct comparisons were more prone to false alarms in identifying targets in target-present lists. The research identifies which model best mitigates cognitive biases and improve decision-making accuracy. The results provide a new evidence-based standard for AFRS interface design, offering specific recommendations for law enforcement.