Research Of Facial Constitution Recognition Based On Deep Neural Networks

Constitution classification is the foundation and core content of constitution research.It extracts the physical characteristics suitable for the group from the complex and varied individual physical phenomena,and finally forms the classification system.At present,the use of deep learning technology in the field of facial constitution classification is still in its infancy,and there are still many problems to be solved.This thesis studies three of them.First of all,the traditional hand-designed features are not strong enough to express facial constitution images.Then,the number of facial constitution data is small.Finally,existing methods don't consider deeply the semantic features of different layers in the neural network.In order to solve the above problems,this thesis has carried out the following research work.(1)In this thesis,we propose a facial constitution recognition algorithm based on deep network feature learning.The algorithm not only enables end-to-end training through convolutional neural network,but also extracts high-level semantic features of the facial image.First,the last full-connected layer of the convolutional neural network is extracted.Secondly,the color features of the facial image are extracted and then merged with the features of the full-connected layer in order to enhance the characterization of facial physique images.Experiments on the facial constitutional dataset demonstrate the effectiveness of the proposed algorithm.(2)In this thesis,we propose an algorithm for constitution classification through transfer learning and integrated learning.Firstly,the Constitution Net was constructed by fine-tuning the Dense Net-169 model on the facial physique dataset.Secondly,we integrate the Constitution Net with VGG-16,Inception v3 and Dense Net-121 to classify the input face image in order to further improve the accuracy of classification.The experimental results on the facial physique dataset demonstrate the effectiveness of the proposed algorithm.(3)In this thesis,we propose a constitution classification algorithm based on multi-level and multi-scale features aggregation.First,the VGG-CI and NASNet Mobile networks were constructed by fine-tuning the VGG-16 and NASNet Mobile models on the facial physique dataset separately.Second,the features of different layers in the VGG-CI network are used and then these features are reduced by feature dimension reduction and merged with the full-connected layer features to enhance the feature representation of facial physique images.Third,the previous layer features of the global average pooling layer in the NASNet Mobile network are extracted.Similarly,these features are dimensionally reduced and then are aggregated with the fused features in VGG-CI network,further enhancing the characterization capabilities of facial physique images.Experiments on facial physique data demonstrate the effectiveness of the proposed algorithm.