Large scale face recognition based on multilevel classification
基于多级分类的大类别人脸识别研究

This paper presented a novel multilevel classification method.In practical face recognition system,the number of the classes is usually large.It caused the correct recognition rate not satisfied if only using one features classified the all probe samples one times.First,marked 20 candidate classes for each probe sample by a two dimensional projection approach.Se-cond,for each probe sample,employed 20 class's candidate training samples to calculate the optimal discriminant vectors,then projected the probe sample and the candidate samples into the vectors and classified the probe sample.The proposed algorithm was evaluated by a 200 persons FERET face database.For elimination the illumination change,adjusted every image's mean and standard deviation to 0.5 and 0.15 respectively.The 10 times experiments average accurate rate directly using combined discriminant analysis is 71.23%,while the average rate of multilevel classification is 83.75%.