Research And Implementation Of Group-level Emotion Classification Method

With the diversified development of human group activities,the identification of group-level emotions has gradually become an emerging research hotspot,which has great value in public security,public culture and other fields.Traditional methods can only collect feedback information on group-level emotions through questionnaire surveys,but this requires a lot of manpower and material resources.With the development of deep learning,emotion recognition technology based on computer vision has made great progress.The category of group-level emotions(negative,neutral,positive)can often reflect the public safety situation to a certain extent.For examples,people at a wedding banquet are very happy,which means that group-level emotions here are positive.And everyone in the library is concentrating on reading books,which means that group-level emotions here are neutral,and people at the fire scene are panic,which means that group-level emotions here are negative.If we know group-level emotions in public places,we can deal with emergencies in time.How to effectively predict and classify group-level emotions in real time is an important content of this thesis.In order to solve the problem that traditional methods are extremely labor-intensive,this thesis combines deep convolutional neural networks,compares different network models and training methods,uses pre-training,migration training and other methods to optimize the training model,and group-level emotions are divided into three categories:negative,neutral,and positive.The classification accuracy rate reaches 76%,and a sentiment group-level emotion classification system is implemented based on the classification model.The specific content is as follows:A method of group-level emotion classification based on SeDenseNet(Squeeze-and-ex citation densely connected convolutional networks)is proposed.First,it analyzes the advantages and disadvantages of the DenseNet(Densely connected convolutional net works)network from a theoretical perspective,and then explains how SeNet(Squeez e-and-excitation networks)can improve it to form a higher-performance SeDenseNet.Finally,through experiments,the performance of DenseNet and SeDenseNet on the problem of group-level emotion classification is compared,which proved the superio rity of this method.Based on the proposed group-level emotion classification method,the core algorithm model is improved,and a group-level emotion classification algorithm based on the dual feature fusion training method is proposed.Since the training sample used contains two labels of group-level emotion category and cohesion index and both of which are related to people's inner activities,it is guessed that there is a certain connection between the group-level emotion category and the cohesion index that promotes the accuracy of the training results.Then the group-level emotion category and the cohesion index were fused together for training.It was found that the fusion training method improved the results of group-level emotion classification and cohesion index prediction,which proved the guess.We designed and implemented a group-level emotion classification system based on the improved algorithm model.Vlc was used to read the camera video stream,Opencv was used to split the frame,Libtorch was used to loads the trained model to classify the video frame and qt was used to make the interface.The three parts of the frame reading,result prediction and interface refreshing was put into different threads for processing.Data was transferred between each thread by the frame buffer queue,which effectively reduced time delay of camera video and improved the forward propagation speed of the model.It achieved the purpose of real-time prediction of group-level emotions,thereby public safety conditions can be obtained to a certain extent and can be reacted in time.