Study On Large-Scale Dataset And Multimodal Image Fusion Methods In Face Recognition

After more than 40 years'development, great progress in AFR (Automatically Face Recognition) technology has enabled its commercial application. However, several practical problems keep AFR far away from mature, especially sensitivity to outer conditions, lower Error Recognition Rate (ERR) for large scale data set, etc. This thesis aims at analyzing these problems and proposing corresponding solutions.Effect of the dataset size on linear and nonlinear methods is firstly discussed. For most face recognition methods, training and self-learning is necessary, where the size including the number of classes and picture numbers in each class should be defined firstly. Changes in the number of classes and picture numbers in each class bring effect to the linear and nonlinear methods. It was found that, CRR with some linear recognition methods decrease greatly after the scale of data set reaching to some extend, while nonlinear method is stable. The essential reason why linear method is disabled lies on overlarge distance within classes and overclose distance between classes. In conclusion, the number of classes and the picture number in each class can be defined according to the distance within classes and distance between classes, and linear methods will have good recognition effect in this case. When distance within classes is overlarge and distance between classes is overclose, the nonlinear method can be used.Decomposition is another method except linear method to solve the critical problem of lower CRR in case of large scale dataset. Two methods for decomposition are proposed, i.e. the decomposition method of face data sets based on clustering and the filter of gallery set based on Cumulative Match Curve (CMC). Further more, the two methods were integrated into a new method– Clustering-CMC. The experiments showed that Clustering-CMC improved CRR, compared with using linear face recognition.For pursuing lower ERR, Multimodal Face Recognition (MFR) based on Visual Image and Infrared Image was studied. MFR integrates information from images of two types by using Image Fusion, and the features of images are more special for classifying. CRR can be improved, compared with images of single type. Three fusion...