| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 97 === For face recognition system, the rejection mechanism is usually ignored, i.e. if the testing person is not in the training set (where we call it “imposter” here), no matter how outstanding the face recognition system is, the imposter won’t be able to be recognized with right label and it will cause high false alarm rate. In order to solve this problem, we propose a rejection mechanism and thus the recognition system incorporating the rejection mechanism not only can classify each test data but also has the ability to reject the imposter data. Moreover, the Local Margin-Enhanced Space is proposed in order to design the rejection mechanism feasibly, where not only the local discriminant data structure can be preserved but the local margin of each data is enhanced based on the k-nearest neighbor classification rule. Hence, the statistic rejection mechanism can be designed by modeling the acceptance and rejection likelihood probabilities according to the distance of each data and its corresponding nearest neighbor in the LME space. Finally, the performance of the proposed system is evaluated using the challenging databases. The results not only demonstrate the system to recognize the single image or image sets (multiple faces) with a high degree of accuracy, but also perform a promising result with 81% rejection rate.
|
|---|
