Electroencephalography (EEG) provides high temporal resolution and noninvasiveness for a range of practical applications, including emotion recognition. However, inherent variability across subjects poses significant challenges to model generalizability. In this study, we systematically evaluated twelve approaches by combining four domain generalization (DG) techniques, Deep CORAL, GroupDRO, VREx, and DANN, with three representative deep learning architectures (ShallowFBCSPNet, EEGNet, and TSception) to enable improved subject-independent EEG-based emotion recognition. The performances of the DG-integrated deep learning models were quantitatively evaluated using two emotional EEG datasets collected by the authors. Data from each subject were treated as distinct domains in each model. Binary classification tasks were conducted to identify the valence or arousal state of each participant based on a ten-fold cross-validation strategy. The results indicated that the application of DG methods consistently enhanced classification accuracy across datasets. In one dataset, TSception combined with VREx achieved the highest performance for both valence and arousal classifications. In the other dataset, TSception with VREx still yielded the highest valence classification accuracy, while TSception combined with GroupDRO showed the best arousal classification performance among the twelve models, slightly outperforming TSception with VREx. These findings underscore the potential of DG approaches to mitigate distributional shifts caused by intersubject and intersession variabilities to implement robust subject-independent EEG-based emotion recognition systems.