Study On Face Detection And Face Attribute Recognition Based On Convolutional Neural Network

Face detection and face attribute recognition refer to detecting all faces in a picture and analyzing facial features of each face,such as identifying the age,gender,race,posture,expression and other face attributes.Among them,face detection is the basis of face attribute recognition,and face attribute recognition can help computers better understand face images,and has huge application scenarios in video surveillance,image retrieval,advertisement placement,human-computer interaction and so on.This paper mainly studies the face detection problem and the three face attribute recognition problems of head pose estimation,gender recognition and age estimation.In terms of face detection and head pose estimation,due to the excellent performance of convolutional neural networks in the field of image classification,more and more scholars use it to solve problems such as face detection and head pose estimation.At present,the joint framework of face detection and head pose estimation based on convolutional neural network can be divided into two categories:one is to first use the face detection network to perform face detection on the input image,and then feed the detected face frame into the head pose estimation network to estimate the pose of each face;the other is to use region proposal algorithm based on image feature descriptors to analyze the input picture and generate a large number of candidate regions,and then sequentially feed each candidate region into the face classification and head pose estimation network based on multi-task learning.It is judged according to the classification threshold whether each region contains a face,and the posture of each face is estimated.In the first method,since the two networks are independent of each other,the definition of the face area is often different.If the face detection output does not match the face area required by the attitude estimation network well,the accuracy of the pose estimation is greatly affected.The second method generates a large number of candidate regions due to the region proposal algorithm,so that the subsequent multi-task network needs to perform a large amount of calculation.And the reliability of face classification based on only one convolutional neural network is not high,which results in high false detection rate,which greatly affects the detection rate of face detection.Aiming at the problems of the above two methods,this paper proposes a joint framework of face detection and head pose estimation based on cascaded multi-task convolution neural network.The framework is made up of three convolutional neural networks of different depths.Each convolutional neural network is a multi-task network that can solve three subtasks:face classification,bounding box regression and head pose estimation.Since face detection and head pose estimation are combined,the problem of face region mismatch in the first method is solved.The three convolutional neural networks work together in a cascade manner,which solves the problem of high false detection rate caused by the second method relying solely on one network for decision making,and achieves strong real-time performance.Three networks were trained using the Wider Face dataset and the AFLW dataset to generate millions of training samples.Face detection performance is evaluated on the face detection dataset FDDB.Head pose estimation performance is evaluated on the head pose estimation dataset AFW.The proposed framework achieves good results in both datasets.In terms of gender recognition and age estimation,similar to other face attribute recognition problems,most research work only focuses on the identification of a single attribute,and lacks a framework that can simultaneously solve gender identification and age estimation.Aiming at this problem,this paper proposes a gender recognition and age estimation network based on multi-task learning,mainly by improving the classification model that achieves excellent performance in the field of image classification,using the convolution feature extraction layer of these models for preliminary feature extraction,and then adding multi-task learning layers.The back propagation algorithm guides the weight update of the network,and finally makes the network suitable for gender recognition and age estimation.The designed network is trained and tested using the public dataset Audience,which includes both age and gender annotations.Good results are achieved in both gender recognition and age estimation.Finally,this paper combines the joint framework of face detection and head pose estimation and the joint framework of gender recognition and age estimation to form an end-to-end face detection and face attribute recognition system.The test is carried out in the actual scene,and the experimental results prove the good performance of the proposed method in various extreme scenarios.