Research On Facial Attractiveness Analysis Based On Machine Learning

Automatic facial attractiveness analysis is to identify automatically human being's beauty using computer vision, pattern recognition, cognitive science and psychology. The facial attractiveness analysis provides technical support to applications such as medical, artistic creation and social networking. Although the beauty of human face images can be easily understood and judged by people, it is hard for the computer since the face image is just a collection of image pixels. How to make the computer to 'learn'a common beauty standard smartly has become an important research direction.Based on the existing facial attractiveness analysis technologies, we develop further research on facial attractiveness analysis based on machine learning. Specially, the main contributions of this dissertation are as follows:①To enhance the intrinsic description for facial attractiveness, the integrated subspace method on the subspaces of PCA and Generalized Low Rank Approximation Matrix (GLRAM) is proposed. Thus, their individual characteristics in characterizing the global and local intrinsic description of facial attractiveness can be collaboratively boosted. In addition, the Gaussian Field model (GF) is applied to reflect the geometry structure in sample space, and better classification results are obtained.②In the facial attractiveness analysis based on geometry property, a new strategy is presented by combining individual local geometry features. Adaboost ensemble learning mechanism is applied to achieve effective integration of features. Besides, Memory Based Dynamically Weighted Kernel Density Estimation (MDWKDE) is proposed to construct weak classifiers and improve the performance of weak classifiers' ensemble learning.③Laplace-constrained shared subspace algorithm is proposed to fully exploit the intrinsic description for facial attractiveness. The appearance information and geometry feature of face images are projected into a shared subspace and generate consistent intrinsic description. Laplacian operator is applied to constrain the geometry structure between the shared subspace and the original subspaces. As a result, the classification performance can be improved. Besides, Online Laplace-constrained Multi-Output Regularized Feature Projection is presented to solve online sample learning problems.