Semantic access of frontal face images: the expression-invariant problem

Semantic queries to a database of images are more desirable than low-level feature queries, because they facilitate the user's task. One such approach is object-related image retrieval. In the context of face images, it is of interest to retrieve images based on people's names and facial expressions. However, when images of the database are allowed to appear at different facial expressions, the face recognition approach encounters the expression-invariant problem, i.e. how to robustly identify a person's face for which its learning and testing face images differ in facial expression. This paper presents a new local, probabilistic approach that accounts for this (as well as other previous studied) difficulty.