Bio-inspired Deep Attribute Learning Towards Facial Aesthetic Prediction

Automatic facial aesthetic analysis is to identify a technology that make the computer simulate the human being's aesthetic analysis mechanism and continue to “learning” the method of human aesthetic.In recent years,with the increasing demand of the facial aesthetic analysis technology in cosmetics,marriage,recruitment,multimedia and other industries,computational prediction of facial aesthetics has attracted ever-increasing research focus.The traditional automatic facial aesthetic analysis technology makes an effort to find features that can affect the facial aesthetics and the high precision classification algorithm to complete the classification and evaluation of different faces.However,these methods can not clearly explain how the human visual system and the brain select and deal with these features in the process of analyzing the effective expression of image information,and lack scientific methods to statistically and verify the rationality.In order to simulate the human aesthetic process as much as possible and get better classification results,this paper proposes a method of facial aesthetics prediction based on middle-level features and deep learning.This method can not only improve the image feature extraction algorithm performance but also improve the accuracy of the classification algorithm by using approaches based on middle-level features representation and model fusion methods,then build a facial aesthetic analysis framework,The specific work is as follows:Firstly,a feature extraction method based on image middle-level features is proposed.Compared with the low-level features,this method is able to express the extracted image features effectively because of its better understanding of the image semantics,and also performs well in the classification effect.This method firstly extracts the facial aesthetic region and the visual significance weight matrix through eye tracking experiment.Then we calculate the Bio-FAO and verify the effectiveness of the Bio-FAO through the "Global + local" visual attention mechanism.Finally,we extract the middle-level features from the Bio-FAO.The purpose of this method is to obtain the image middle-level features representation,compared with the method of image low-level representation,this method has better ability to express it because of its better understanding of image semantics.it also have a better classification ability.The experimental results show that our method has a better effect in facial expression.Secondly,a method of calculating the facial aesthetics label based on True Skill algorithm is proposed.Firstly,a sample set composed of the same size face image is established.The relative sorting results of the facial aesthetics are obtained by the aesthetic sort experiment,we then calculate the absolute aesthetics score through the Trueskill algorithm,finally calculate the facial aesthetics label through the weighted average algorithm.Thirdly,a facial aesthetic analysis framework based on deep learning and model fusion is proposed.Firstly,we trained a set of bio-inspired deep attribute detectors by cutting-edge convolutional neural networks and get the the probability values of the training samples,then the probability values of each training samples in different detectors are merged into the eigenvectors of the samples,Finally,we combine with facial aesthetics label training the facial aesthetics prediction model.This framework improves the accuracy of the classification algorithm.The experimental results show that the method based on deep learning and model fusion not only has higher classification accuracy than a single bio-inspired deep attribute detector,but also shows good results compared with other feature extraction methods.