Research And Implementation Of Face Attribute Recognition Based On Deep Learning

With the continuous development of computer science and technology,the recognition technology based on face image has become a hot research issue in the field of computer vision.Age estimation technology and expression recognition technology based on face images have a wide range of application prospects in daily life and commercial market,which can be used in permissions control,personalized service,video recommendation and so on.However,some methods of face analysis are often carried out in the ideal state,while in the real scene,they are often affected by many environmental factors.The most common application scenario is face analysis in video surveillance scene.Therefore,it is a difficult task to solve the problems of face occlusion,face scale change and face dynamic change in video surveillance scene.This paper is mainly based on deep learning technology to improve the relevant content of face analysis for the above problems,mainly including face detection,age estimation and facial expression recognition.The specific work of this paper is as follows:(1)An improved MTCNN face detection algorithm is proposed to solve the problem of face occlusion in crowded situations.First,we build the MTCNN network structure and generate the required training samples.Then,each level of network is trained one by one.In the process of training sample generation and network training,in order to reduce the predicted number of candidate boxes,the candidate boxes with large overlapping degree are usually deleted.This step usually uses a Non Maximum Suppression algorithm.This paper improves the Non Maximum Suppression algorithm,optimizes the confidence reset function,resets the confidence score of candidate boxes that overlap with the detection box in a decaying manner.At the same time,the separable convolution is used instead of the traditional convolution to improve the network structure.The experiments show that the improved MTCNN face detection algorithm improves the detection rate and reduces the number of parameters in the case of face occlusion.(2)In order to solve the problem that the estimation of face age in different scales of the same face will have certain deviation,a multi-scale multi-stage face age estimation algorithm is proposed.Firstly,the pre-processed samples were processed with bilinear interpolation to obtain multi-scale samples.Then the samples of different scales are separately input to the feature extraction network and feature fusion is carried out in stages.At last,the predicted age of multiple stages was added to obtain the final age regression prediction.Experiments show that the input of multi-scale samples can improve the accuracy of age estimation.The MAE values in the IMDB and WIKI datasets are 6.44 and 6.31,respectively.The multi-stage regression prediction makes the size of the network model only 0.51 MB.So,the method in this paper can accurately estimate the age of human face under the condition of small model and different scales.(3)Aiming at the differences of facial expressions extracted from different modalities,a video facial expression recognition algorithm based on modal fusion was proposed.First,a modal fusion algorithm framework is designed.Then,C3 D and CNN-LSTM methods are used to extract the spatialtemporal features of the image sequence.Finally,two different fusion strategies are studied: feature fusion and result fusion.Experimental results show that the recognition rate of modal fusion is higher than single mode,the recognition rate of three-modal fusion is higher than that of two-modal fusion,and the recognition rate of feature fusion is higher than that of result fusion.