Similarity and fuzziness in visual information management

We compare several possible approaches to the measurement of similarity and differences in visual information management systems. For many visual information retrieval applications, it is important to order the data according to a scale as close as possible to the human judgement. Traditional techniques based on the measurement of Euclidean distance between points in feature space prove often unsatisfactory. Data about the measurement of similarity can be derived from the psychological literature. F. Attneave (1950) performed a series of experiments to measure how the perception of similarity changed when a number of features of the stimulus were changed continuously. His findings definitely rejected the hypothesis that the human perceptual system may compute a Euclidean distance between the two stimuli it compares. He found a strong nonlinearity that suggested that the same variation in the features was weighted more when the two stimuli were very similar than when the two stimuli were wide apart. Also, he supported experimentally an hypothesis due to Householder that distances along independent perceptual axes were summed, as in the “city block” distance