Evaluation Methodology And Experiments For Face Recognition Systems

In the recent years, face recognition has been one of the key research points in the image analysis and understanding area. The development of face recognition technology is due to three main factors: the development of face recognition algorithms, the availability of large-scale facial image databases, and the methodology for the evaluation of face recognition algorithms and systems [10]. Since face recognition technology bases itself on large amount of image data and the existing face databases are deficient in their scale and variation conditions, the construction of a new large-scale face database will greatly promote the development of face recognition technology. Evaluating the face recognition systems tests the performance of algorithms and systems on one hand, on the other hand, it also has the function of inspecting the status of the whole technology and pointing out some valuable research interests. In this thesis, we focused on the design of facial image database and evaluation methods for face recognition systems, and made the following contributions:1) Designed and constructed CAS-PEAL large-scale face image databaseThe CAS-PEAL face database consists of 99,450 facial images of 1,040 Chinese individuals with four principal variations of Pose, Expression, Accessory and Lighting, as well as three other variations of Background, Distance and Aging. This thesis introduces the detail information about the CAS-PEAL including construction environment, structure of the image files, and sample images in the database. In addition, the thesis describes the release version of the CAS-PEAL: CAS-PEAL-R1.2) Discussed the methodology for evaluating face recognition algorithms or systemsWhether the goals of face recognition evaluation work can be achieved depends on two main factors: evaluation methods and evaluation criterions. This thesis introduces and discusses the principles of evaluation design, as well as the main evaluation methods and the relevant criterions.3) Evaluated several baseline algorithms and systems based on CAS-PEALWe realized two baseline face recognition algorithms (Eigenface and Correlation) and two outstanding business systems. Based on the CAS-PEAL database, we tested the performance of each system. This thesis reports the evaluation process of the two baseline algorithms and two systems in detail and analyzes the evaluation results. Keywords: face recognition, face image database, performance evaluation...