Research On Emotion Recognition Based On PCANet

Emotional human-computer interaction is an important component of artificial intelligence in areas such as driver emotion detection,emotional assistance robots,and studying state monitoring.Emotion recognition research has emerged as a hot spot in the field of artificial intelligence today.Emotion recognition based on physiological signals is more authentic and reliable than surface signals such as facial expressions and speech.The majority of studies on emotion recognition based on physiological signals use electroencephalogram(EEG)signals or fusion physiological signals as the research object,and studies on emotion recognition based on electrocardiogram(ECG)signals are fewer.EEG-based emotion recognition research has advanced in recent years,but collecting EEG signals requires electrodes to be attached to the head,which has poor experience and is not suitable for daily emotional monitoring.The development of portable ECG signal acquisition equipment(such as a smart watch that obtains real-time ECG signals)allows emotion recognition based on ECG signals to be applied to daily emotional monitoring.However,traditional machine learning's emotional representation of shallow features is insufficient,and the emotional information contained within physiological signals cannot be fully mined;deep learning extraction of deep features has issues such as time-consuming,dimension disaster,and feature redundancy.In light of the aforementioned issues,this paper primarily conducts the following studies:A PCANet-Based Mutual Information Feature Selection Network(PCAMI-Net)is built on the Principal Component Analysis network(PCANet).PCANet extracts depth features using Principal Component Analysis(PCA)convolutional layers and has achieved good results in pattern recognition,image processing,and other fields.The PCANet-generated high-dimensional deep features have feature redundancy,and the recognition accuracy needs to be improved.Based on PCANet,this paper proposes an emotion recognition model named PCAMI-Net based on ECG signals.PCAMI-Net removes redundant features contained in high-dimensional features on the basis of fully mining deep emotional features to solve the problem of feature redundancy.PCAMI-Net selects features that are more closely related to emotions based on mutual information between features and emotion categories to address the problem of improving recognition accuracy.The recognition accuracy can be improved in this way.The AMIGOS dataset is used for the emotion recognition research.The accuracy rates in the arousal and valence dimensions in the user-dependent experiment were 72.5% and 60.3%,respectively;in the user-dependent balance experiment,the accuracy rates were 78.3% and 82.8%,respectively.The emotion recognition effects of PCANet and PCAMI-Net with four classifiers were also compared and analyzed.PCAMI-Net has a better emotion recognition effect and a shorter training time than PCANet,according to comparative experimental results;in user-independent experiments,PCAMI-Net's arousal and valence classification accuracy rates reached 57.2% and 49.5%,respectively.A PCANet-based attention mechanism emotion recognition model is constructed.PCANet extracts high-dimensional features using PCA convolution,and the weight coefficients of each feature are consistent,making it difficult to highlight the role of useful features.PCAMI-Net chooses features.Although it can improve feature quality,it is also simple to filter out some valid data.In this paper,an attention mechanism emotion recognition model based on PCANet is proposed,which retains all features and improves classification performance by changing the weight of features.A Bidirectional Long Short-Term Memory Network(Bi LSTM)layer is used to replace the encoding layer in order to retain the original convolution features.The Bi LSTM layer employs a unique gate structure to retain valid information while discarding redundant information.The attention mechanism is introduced in order to intelligently focus on the emotional information part.The AMIGOS dataset is used for the emotion recognition research.In user-independent experiments,arousal and valence classification accuracy rates reached 58.1% and 50.9%,respectively.The results show that the model outperforms PCANet and PCAMI-Net.Furthermore,the emotion recognition effects of four deep models were compared and analyzed,and the model's effectiveness was confirmed.