ES-RU: an entropy based rule to select representative templates in face surveillance

• Published in: Multimedia tools and applications Vol. 73; no. 1; pp. 109 - 128
• Main Authors: De Marsico, Maria, Nappi, Michele, Riccio, Daniel
• Format: Journal Article
• Language: English
• Published: Boston Springer US 01.11.2014; Springer Nature B.V
• Subjects: Algorithms; Analysis; Biometric identification; Biometrics; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Diversification; Entropy; Face; Facial recognition technology; Galleries; Identification; Indexing; Modules; Multimedia Information Systems; Neural networks; Pattern recognition systems; Special Purpose and Application-Based Systems; Studies; Surveillance; Tasks; Biometrics; Video surveillance; Entropy; Face indexing
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-012-1279-6

ES-RU is a system for video sequence indexing. Video frames are annotated according to the identities of appearing subjects. The system architecture is designed by distributing the different processing steps across dedicated modules. These modules interact with each other to accomplish the final task. Such modularity is also designed to allow a high system flexibility, because it is possible to independently substitute each component with a different one performing the same task using a different method. As an example, face detection is presently performed by Viola–Jones algorithm, but the corresponding module might be substituted by one exploiting neural networks or support vector machines (which are actually more computationally demanding). In detail, ES-RU implements both face location and analysis, and an algorithm to select the most representative templates for the selected identities. The novelty of the algorithm for template analysis and selection relies on the proposed use of the concept of entropy. This concept is the base of most techniques that exploit relative entropy to estimate the degree of uniqueness which is assured by a biometric trait, when processed by a Feature Extraction Technique (FET). In this paper, entropy is introduced as a tool to evaluate the contribution of each sample in guaranteeing a suitable diversification of the templates that make up the gallery of a relevant subject. Video-surveillance activities cause to gather a huge amount of templates to be used for tracking and re-identifying subjects. However, most of these templates are not informative enough to be useful. The aim of our approach is to provide an effective technique to keep only the most “representative” of them, i.e. those that provide a sufficient level of diversification. This allows faster processing (less comparisons) and better results (it is possible to recognize a subject under different conditions). ES-RU was tested on six video clips and on a subset of the SCFace database to assess its performances.