Research On Automatic Extraction And Classification Of Facial Components

Computer simulation of face portraits plays an important role in public security work.Through the description of facial features by victims or witnesses,the appearance of criminals can be detailed through the components in the face component library.However,the facial component library used for computer simulation of face portraits is not rich enough and has a limited scope of application.In order to expand the facial component library quickly and improve the efficiency of case processing,this paper deeply studies the automatic extraction and classification of facial components.The specific work and innovation are as follows:In terms of facial components extraction,MTCNN neural network was built on the TensorFlow framework.Based on the CelebA dataset,the faces in 10000 images were manually labeled by using the lableme annotation tool to tag corresponding tags to build dataset need for network training.Finally,95% of the accuracy of face detection is achieved.The four-level neural network was used to obtain facial landmark.Then,two-point extraction method and three-point extraction method were designed for different parts to extract facial parts one by one in the form of rectangular boxes.In the classification of facial components,two methods are proposed.One is based on the feature semantics.Four kinds of facial parts,including eyes,eyebrows,nose and mouth,were analyzed.The corresponding mathematical models were constructed and shape parameters were defined.Then the categories of components were determined according to the value range of different parameters,and a semantic facial component library was established.The other method is based on the convolutional neural network.Using web spider,we got 108806 pictures uploaded to the album by 122 stars of sina weibo,and preprocessed them by face detection.In addition,CAS-PEAL dataset is introduced.Part category labels were added to the faces in these two data sets by manual annotation.For different face components,different VGG networks were used for training,and after several epoches of training,the classification accuracy rate is over 80%.In terms of the design and implementation of the software for automatic extraction and classification of facial components,by analyzing software requirements,writing codes on PyCharm2017,using PyQt5 to design the interface and combining with the PyTorch deep learning framework,the software for automatic extraction and classification of facial components is designed and implemented.It includes three functional modules: "Face detection"," Facial component extraction" and " Facial component classification".