Biometric recognition algorithms spring up in recent years for their widely applications in the guard of national and public security. The expanding market of biometric products is speeding up the development of biometric algorithms. With those miscellaneous biometric algorithms coming into being, what really matters is to find a standardized platform and efficient evaluation method to assess whether the specific biometric algorithm is efficient and robust enough for the practical applications. This thesis elaborates theoretical analysis of evaluation models combined with experiences of implementation of the biometric algorithm evaluation platform. More concretely, the following chapters will cover the construction of database, protocol design, quality evaluation and error and confidence analysis. The main contributions of our work are as follows: 1. We design and develop an evaluation platform, which contains four modules:database module, testing protocol module, performance analysis module and prediction module. 2. In the proposed database module, we improve software of NFIQ (which is developed by NIST) to evaluate quality of fingerprint databases. 3. In the testing protocol module, three kinds of strategies are proposed: the novel bootstrap based matching strategy, the improved NIST matching strategy and user-defined strategy. 4. In the analysis of algorithm performance module, we adopt the subset bootstrap based method to calculate confidence interval of error rates which is testified to achieve good results. 5. With regard to evaluation of database quality, we propose a brand new hierarchical model. This work includes the following characteristics: the quality is evaluated based on hierarchical influencing factors, local and global factors; the quality is determined quantitatively according to distribution of genuine matching scores; a subset frequency method is proposed to model relations between distribution of genuine matching scores and quality of corresponding images; the model can be extended to all modalities; quality evaluation at each level in this model can be used independently in applications. 6. We propose a performance prediction model based on Multi-Gaussian distribution assumption and present corresponding experimental results and conclusions. Generally, this thesis focus on the new issues in biometric algorithm evaluation. Firstly, we construct biometric evaluation platform successfully; secondly, we propose novel hierarchical mo...
修改评论