Research Of Facial Attractiveness Computing Based On Deep Convolutional Neural Network

Currently,the features of facial attractiveness computing mainly focus on geometrical features,regardless of the factors including the skin texture?ornament and expression.In order to represent facial attractiveness more effectively,we study the shallow facial attractiveness representations and deep attractiveness representations together,and propose two facial attractiveness classification methods,one is based on face similarity retrieval strategy and the other one is based on facial attractiveness CNN feature and facial geometrical feature.The main work includes:1)Face similarity CNN.We construct a face similarity convolutional neural network based on maxout CNN using two supervised signals,the face classification signal and face verification signal.Besides,we use the tricks of face alignment and multi-region face representations extraction methods.With the help of large number of training samples,our CNN achieves 97.25% face verification precious in LFW test data.2)According the priori condition that similar faces' facial attractiveness are close,we propose attractiveness classification methods based on face similarity retrieval strategy.We use Euclidean distance as distance metric and the face similarity CNN feature as image representation.According the priori condition,we compute the retrieved faces' attractiveness labels' expectation to get the test face's attractiveness.This method achieves a good classification result on test dataset created by us.3)Deep feature and shallow feature fusion.We propose facial attractiveness classification method based on attractiveness CNN feature and geometrical feature pool.The face similarity CNN is finetuned by the supervised signal of attractiveness classification and the geometrical feature pool consists of several traditional geometrical features.We show that the fusion feature of attractiveness CNN feature and geometrical feature pool has powerful representation ability,considering not only the face's coordination but also the high-level attractiveness semantic information.