ResNet-50 based technique for EEG image characterization due to varying environmental stimuli

Abstract

Background and Objective

Emotion is an important factor affecting a person's physical and mental health, but there are few ways to detect a patient's emotions in daily life. Negative emotions not only affect recovery after treatment, but also cause poor health. Current emotion classification research based on EEG image recognition is highly accurate, making the development of an emotion detector feasible. Using emotion data from the SEED, this study trained a detection model using the residual neural network ResNet-50 with a SAM and SE-block double attention mechanism, and used quantitative parameters based on the Russell emotion cycle model to construct a human–computer interactive health detector for emotion recognition in EEG images induced by environmental stimuli.

Methods

Images of 61 environmental scenes were collected and divided into three categories according to the visual characteristics of the environment. Eight volunteers were recruited to collect a total of 488 EEG image data. The trained ResNet-50 model was used to automatically analyze the characteristics of the collected EEG images and classify emotions. The model was compared the support vector machine (SVM), transfer component analysis (TCA), dynamic graph convolutional neural network (DGCNN), and DAN methods.

Results

The accuracy of the ResNet-50 model trained in this study is 85.11% and its variance is 7.91. Through the verification of EEG images induced by environmental stimuli, the results are improved by 2.01% and the variance is reduced by 0.04 compared with the model's training results. The model is more accurate in identifying negative and neutral emotions, indicating that the ResNet-50 architecture better recognizes motions in EEG images induced by environmental stimuli. Compared with other algorithm models, the proposed model has the lowest variance and highest stability. The comparison of various algorithms revealed that environmental scenes with different visual features induce different emotions.

Conclusion

The proposed monitor can collect EEG images of patients induced by environmental stimuli in daily life in real time, automatically analyze and identify emotional characteristics, and provide quantitative parameters and visualization. It not only enables patients to conveniently monitor their emotional state and make timely adjustments, but also assists doctors in clinical diagnosis.