Text Driven Recognition of Multiple Faces in Newspapers

Face recognition is still a hard task when performed on newspaper images, since they often show faces in non-frontal poses, prohibitive lighting conditions, and too poor quality in terms of resolution. In these cases, combining textual information derived from the page articles with visual information proves to be advantageous for improving the recognition performance. In this work, we extract characters’ names from articles and captions to restrict facial recognition to a limited set of candidates. To solve the difficulties derived from having multiple faces in the same image, we also propose a solution that enables a joint assignment of faces to characters’ names. Extensive tests in both ideal and real scenarios confirm the soundness of the proposed approach.