Design And Research Of Face Attraction Template Based On Transfer Learning And Geometric Characteristic Analysis

Face attractiveness is an important proposition of genetic psychology and cognitive psychology.The research results can provide scientific basis for face evolution(and human evolution).Face attractiveness study combined with computer can improve the shortcomings of traditional research methods,such as strong subjectivity and less data.In the research of face attractiveness,it is the primary task to determine the facial template.This paper proposes a new template of face attractiveness based on the transfer learning.Three geometric features that represent the face attractiveness are extracted based on the face template.Experiments show that the template of face attractiveness is reasonable and feasible.The main research contents are as follows:(1)In the localization of face feature points,the method of combining AAM and ASM algorithm to locate key points is analyzed.This paper combines AAM and ASM aiming at the problems that the global contour localization of AAM algorithm and the local organ localization of ASM algorithm are not accurate.ASM is used to locate the facial contour area,and AAM locates the facial internal area,which make the facial feature point localization get better results.(2)The template of face attractiveness based on transfer learning is proposed.The face attractiveness template with 81 points is constructed starting from the basic idea of transfer learning,using the combination algorithm of AAM and ASM,considering the influence of the geometric features of forehead,pupil and nose wing on the face attractiveness.Experiments show that the constructed 81-point template of face attractiveness has good universality on different data sets.(3)The 21-dimensional geometric features,the triangular area features and the four ratio feature sets are extracted to represent face attractiveness based on the constructed 81-point face attractiveness template.The machine learning algorithm is used to evaluate prediction performance of face attractiveness of extracted geometric features.The experimental results show that the combined geometric features are better at representing the face attractiveness.In the SCUT-FBP library,the Pearson correlation coefficient is 12.2% higher than the single 81-point feature.In the self-built data set,the Pearson correlation coefficient is 11.4% higher than the single 81-point feature,which further proves that the constructed template of face attractiveness is feasible in this paper.