基于时间序列成像的多任务学习驱动情感识别OA

To overcome the high computational complexity of EEG-based emotion recognition methods based on fea-ture extraction or time-frequency representations,a multi-task learning-driven method for emotion recognition based on time series imaging(TSI)was proposed.EEG signals were directly transformed into two-dimensional images using Gramian angular field,Markov transition field,and motif difference field.Built upon the ResNet18 architec-ture,a multi-task feature fusion framework was designed to integrate features from the three imaging methods to en-hance emotional feature representation.Experimental results showed that with the DEAP dataset,the proposed method achieved average classification accuracies of 96.51%and 97.22%for binary classification of Valence and Arousal,respectively,and with the AMIGOS dataset,the accuracies reached 98.59%and 99.64%.When extend-ed to four-class and eight-class classification tasks,the proposed method achieved average accuracies of 91.06%and 87.43%with DEAP,and 97.41%and 89.84%with AMIGOS,respectively.These results demonstrated the robustness of the proposed method in EEG-based emotion recognition.