Recognition For Facial Expression Of Infants Based On Convolutional Neural Network

Due to the lack of verbal description ability,0-1 years old infants can only express their emotions and physical conditions through different facial expressions.If the different facial expressions of infants cannot be correctly identified and the information expressed by infants cannot be fed back in time,it will cause a series of adverse effects on the growth and development of infants.Therefore,accurately identifying various expressions of infants is of great significance for scientific parenting and has a wide application prospect.This paper is based on Convolutional Neural Network for face detection and correction and expression recognition of infants.The research work includes the following three points:(1)The infant images database and infant expression images database are established.The appropriate database is the basis for the facial research of infants.This paper constructs a database of infant images containing 7,532 pictures through video capture and image collection.Because the database has a lot of background information and some pictures have facial deflection problems,then through the improved MTCNN processing,the vertical facial images of infants with eyes in a horizontal state are obtained.After the evaluation operation,a database of infant expression images is established,which including 5,600 pictures of calm,happy,cry and sleep facial expressions.(2)The infant face detector is constructed and the infant face correction is realized by precise positioning of the center points of the left and right eyes.Due to the great differences between infant and adult images,the existing face detectors are not ideal for applying directly to infant images.Therefore,this paper focuses on the MTCNN algorithm and constructs an infant face detector based on the infant images database.Through experimental comparison,the feasibility and robustness of the application of MTCNN in infant face detection are verified.Then according to the deficiency of MTCNN in infant face correction,the MTCNN is improved by combining facial features of infants,a new CNN is cascaded to achieve precise positioning of the left and right eye center points.The experimental comparison shows that the improved MTCNN reduces the positioning error of the lefit and right eye center points in different states,and improves the accuracy of positioning,and achieves the effect of accurate correction of infant's face.(3)The BabyNet is constructed to realize the recognition of different facial expressions of infants.Due to the small sample size of the infant expression images database and small categories of expressions,this paper designs a new network called BabyNet based on the structure of AlexNet and VGGNet to study the expression recognition of infants.The experimental comparison shows that BabyNet has more advantages in the recognition of different expressions on the infant expression images database.For the unevenness of the recognition rate of different expression categories in BabyNet,the weighted Softmax cross entropy loss function is used to optimize.The experiment shows that the BabyNet using the weighted Softmax cross entropy loss function achieves the equalization of different category recognition rates.Finally,a simple infant expression recognition system is constructed,which can achieve the effect of automatic face detection and expression recognition of infants in the input images.