Man-Machine Interaction in Human-Face Identification

Man-Machine Interaction in Human-Face Identification By A. J. GOLDSTEIN, L. D. HARMON, and A. B. LESK (Manuscript received September 20, 1971) How well can a computer identify a human jace which is described by a person who is inspecting a photograph? We give an account of an interactive system that takes advantage both of the human's superiority in detecting and describing noteworthy features and of the machine's superiority in making decisions based on accurate knoidedge of population statistics of stored face-features. Experiments using a population of 255 faces and 10 or fewer feature-descriptions showed that the population containing the described individual c.oidd be narrowed down to less than If. percent in 99 percent of all trials. I . I N T R O D U C T I O N In a previous report 1 we described experiments in human-face recognition which were intended to establish a foundation for extended study. Those experiments provided a large body of reliable quantitative data based on 21 feature-descriptions of 255 human faces. These 21-dimensional vectors were shown to be sufficient for accurate individual identification, both by human and by computer search.* The objective, then and now, is to explore new techniques for obtaining accurate recognition of vectors given imprecise component values. Our procedures involve searching through a population of vectors to retrieve one, a "target," whose components best match a searcher's imprecise specification. There are two obvious kinds of such recognition and retrieval, just as in fingerprint-file search.