Multi-face Expression Recognition In Video Based On Convolutional Neural Network

In recent years,computer vision has experienced a big development,facial expression recog-nition is one of the most important technologies.The main research contents of this thesis include two aspects,face detection in video and recognition of facial expression based on CNN.The description of research work is shown as follows:Face detection in videos is a key preprocessing step for following facial expression recog-nition.The detection rate of AdaBoost algorithm is not ideal under complex backgrounds,while the use of skin color to conduct face detection has a good effect but at the same time has a high false detection rate.Therefore,this thesis proposes a method that combining Ad-aBoost algorithm and segmentation of skin color for face detection in videos.First,skin color segmentation is applied to eliminate effect of complex backgrounds and light sources,and then the face candidate regions obtained in the previous step are used as input subwindows of AdaBoost to classify.During the training of Adaboost's classifier,the weight update is combined with the positive and negative sample false detection rates to suppress the over-fitting of the weights in similar areas of the face.According to the experimental results,this method can get a higher face detection rate and also reduces the false detection rate.Due to the variety of face angles in videos and occlusion,traditional facial expression recog-nition algorithms that rely on manual feature extraction are not only complicated but also eas-ily cause feature loss.Therefore,this thesis introduces deep learning into expression recog-nition task.Convolutional neural networks can extract features from images autonomously through neural networks,and have the advantages of local response and weight sharing,which is widely used in tasks such as recognition of facial expression.The pooling algorithm is one of core technologies of CNN.By performing aggregation statistics on the features of the convolutional layer,pooling operation can improve the ability of feature representation and also can decrease feature dimension of the CNN.However,the currently used pooling algorithms still have problems of lack of flexibility and single feature extraction.Because of these disadvantages of these existing pooling algorithms,this thesis proposes an improved adaptive pooling algorithm according to characteristics of deep learning that can use BP algorithm to autonomously adjust parameters.This improved algorithm can continuously optimize parameters in the pooling domain based on the value of loss function during the training process,and finally let the difference between the value of actual result and the value of predicted expression become smallest.According to experimental results based on CK+and FER2013 facial expression databases,compared with existing pooling algorithms,the proposed adaptive pooling algorithm can effectively improve the accuracy of expression recognition.