Probabilistic ISC for Matching Images of Objects Having Individual Difference

摘要

This paper considers image matching for objects with individual differences and deformations, such as the human face, and proposes the concept of probabilistic increment sign correlation (ISC) as a new statistic suited to the purpose. Probabilistic ISC is a statistic based on the probability of occurrence of incremental signs calculated from multiple reference images. Since matching is sought considering only the increase or decrease of the spatial brightness, it is less affected by changes of illumination and other factors. In matching, the variation of the incremental sign produced by changes of the object shape and other factors is represented by a probability, and high matching accuracy is achieved by assigning larger weights to features with smaller variation. The computation cost is as low as that of increment sign correlation, and the method is also suited to hardware implementation. It is possible to set the matching threshold analytically on the basis of statistical properties. In order to verify the effectiveness of the proposed method, an experiment was performed to detect a face from 2420 images, and higher detection accuracy was obtained than by methods based on correlation, such as normalized correlation and increment sign correlation, or the subspace method.