Research On Dimensional Emotion Recognition Based On Deep Learning

The dimensional emotion recognition task enables computers to automatically recognise human mental states by recognising a given image,video or voice message to quantitatively describe the human mental state,which enables computers to provide a more meaningful service to humans.Traditional methods using manual features lack generalisation modelling capabilities for large scale data,while deep neural network based methods have significant advantages in feature extraction and high dimensional feature processing,hence dimensional emotion recognition tasks are now usually based on neural networks.The main research contents of this paper are as follows:(1)A dimensional sentiment recognition method based on Res Net and optimized RNN is proposed.The optimized RNN is divided into LSTM and GRU,and feature extraction is completed using cascade structure to avoid the model complexity and pre-training difficulties caused by increasing dimensionality.Experiments are conducted on the Aff-Wild2 and AVEC2016 datasets to verify the validity of the proposed method,including a comparative analysis of the two methods.(2)A dimensional emotion recognition method based on Res Net and optimized TCN is proposed,in which a temporal attention module is added to the TCN network to capture the relationship of feature points within the sequence,and proposes a dual dilated convolutional structure to solve the problems of the small perceptual field of shallow network and the inability of deep network to perceive local information of TCN.The experimental results show that the method has good performance in emotion recognition.(3)A multi-task and multi-modal dimensional emotion recognition algorithm is proposed to exploit the intrinsic correlation between different emotion representations to assist the dimensional emotion recognition task with an emotion classification task.And a feature fusion method based on Leader-Follower idea and attention mechanism is used to fuse audio features as well as video features.The experimental results show that the method achieves excellent recognition results.The network model proposed in the paper has been tested on Aff-Wild2 and AVEC2016 with average recognition accuracies of 0.624 and 0.701 respectively,and compared with other mainstream dimensional emotion recognition models,the model in the paper is found to have higher recognition accuracy.