Skip to Main Content
One of the main challenges in face recognition is represented by pose and illumination variations that drastically affect the recognition performance. This paper presents a new approach for face recognition based on Regularized-Labeled Linear Discriminant Analysis (RL-LDA) learning from 3D models. In the training stage, 3D face information is exploited to generate a large number of 2D virtual images with varying pose and illumination, and these images are grouped into different labeled subsets in a supervised manner. Labeled Linear Discriminant Analysis (L-LDA) is operated on each subsets subsequently. On this basis, eigenspectrum analysis is implemented to regularize the extracted L-LDA features. Recognition is accomplished by calculating RL-LDA features, and achieved a recognition rate of 98.4% on WHU-3D-2D database.