A Research Of Infant Expression Recognition

With the development of artificial intelligence,the research in human-centered intelligent recognition has become a hot trend in the computer field,and the research results have been gradually applied in many applications,such as face recognition,facial expression recognition,object detection and so on.Facial expression recognition as a way to detect human emotions naturally has drawn a lot of researcher's attention,but most of the current research focus on facial expression recognition in adults.As the most important stage of human beings,the infant's language expression ability is not yet perfect,the judgment of the baby's mood and state can only be through facial expressions and body movements.Therefore,if the intelligent recognition of infant's facial expression can be realized,it will bring great convenience in many aspects,such as baby care,nurturing,health detection and so on.In this thesis,taking infant facial expression recognition as the subject,the main research contents focus on the following points:Firstly,an infant facial expression database suitable for this classification study is established.At present,most of the facial expression databases at home and abroad are adult facial images or sequences,and the facial features of infants are quite different from those of adults.Therefore,this thesis uses the way of collecting pictures from the Internet to preprocess and enhance the collected pictures,and finally establishes a infant expression database suitable for this study,which can be divided into four categories: happy,crying,quiet and sleeping.Secondly,the existing LBP(Local Binary Pattern)operators are improved.In view of the fact that the expression data in this thesis are from the Internet and have the particularity of complex illumination and noise,the existing LBP operators are improved.Experiments show that compared with the traditional LBP operator,this operator can highlight the facial features of infants,and has achieved better robustness to light and noise.Thirdly,a deep learning method is used to recognize infant's facial expressions,and a two-channel model combining shallow network and deep network is proposed.In this thesis,the research in infant expression recognition includes three parts: shallow network,deep network and two-channel network combined with two models.The shallow network part introduces LBP feature map,which improves the performance of shallow network by nearly 10 percentage points,and the recognition rate is up to 82.6%.In the depth network part,the classical image recognition network VGG is fine-tuned to make it suitable for infant expression recognition in this paper,and the accuracy is up to 87.3%.In the two-channel network model,the first two networks are fused,and the accuracy is up to 91.6%.Finally,a new type of second-order convolution neural network is adopted for infant expression recognition.Because convolutional neural networks only extract the firstorder features of images,there are more and more applications of second-order features of images in depth learning in recent years,and some studies show that its performance is better than that of first-order features.Therefore,the second-order convolution neural network is further used to study infant expression recognition,and the model is further improved in view of the existing problems in the application process.