| With the rapid development of social economy,the problem of population aging is becoming increasingly prominent.How to use artificial intelligence to deal with the problem of old-age care has become a hot topic in social research.The research on facial expressions and emotions for the elderly can realize the psychological and emotional detection of the elderly.It has certain help to empty nest families and has great application prospects in medical treatment.Aiming at the problems related to the facial expression recognition system,the thesis starts from three aspects: the establishment of the elderly facial expression database,the feature extraction of traditional machine learning algorithms and deep learning algorithms,and facial expression recognition.According to the current situation of the lack of expression database for the elderly and the lack of research on the emotion recognition of the elderly,this article collects images in many TV dramas about the elderly as the protagonist,and establishes its own expression database for the elderly according to the specifications for establishing the face image library.In order to eliminate the problems of background interference and different sizes and positions of faces,preprocessing such as face alignment,size normalization and gamma correction was performed on the image.Gabor wavelet has strong spatial position and direction selectivity,and can well extract the features of the target image in different spatial positions,frequencies and directions.The Hog feature is very robust to change in the environment and can effectively extract image texture features.Therefore,this paper proposes an expression recognition method based on the T region of the face and an expression recognition method that combines Gabor and Hog features of the expression image.In view of the problem of redundancy in the obtained dimensions,the downsampling and PCA dimensionality reduction methods were used in the experiment to reduce the dimensions of the image.The features after dimensionality reduction were classified under different kernel functions of SVM.The experiment showed that the T-zone feature of the face can not only effectively reduce the learning time,but also improve the accuracy of facial recognition to a certain extent;and the experimental results of the fusion feature were also improved compared to before fusion.The accuracy of facial recognition was also improved,which verifies the fusion of Gabor and Hog Thevalidity and rationality of features for facial expression recognition.Deep learning has a strong feature learning ability.When training a recognition model,you can directly input the original image without performing complex preprocessing on the input image.The different network structures of deep learning can not only reduce the amount of calculation,but also effectively achieve image dimensionality reduction.In this paper,the model of convolutional neural network was used to recognize the expression of the elderly.In view of the fact that deep learning models were often prone to overfitting when processing small sample data sets,this paper proposed an expression recognition method combining convolutional neural network and support vector machine.CNN was used to extract facial features first,and then SVM was used for classification.Experiments showed that the expression recognition method combined with convolutional neural network and support vector machine not only trains faster in elderly facial expression classification,but also has higher recognition accuracy.It can be seen from the comprehensive experimental results that the happy and angry expression features of the elderly are more obvious,feature information is easier to extract,and the recognition accuracy rate is higher.Neutral expressions are most easily mistaken for other expressions.Comparing the experimental results of the traditional machine learning method and the deep learning method in the expression library of the elderly,it can be seen that the advantages of deep learning in learning expression features not only can effectively improve the recognition rate,but also have a fast learning efficiency.Figure [39] table [15] reference [82]... |
