To fully extract the frequency information and spatial topology information of multi-channel EEG signals, this paper introduces an EEG emotion recognition model——CBAM-CapsNet, which utilizes a capsule network with a hybrid attention module. Firstly, EEG signals from different frequency bands are acquired, and their differential entropy features are extracted. Secondly, these features are mapped into a three-dimensional compact feature matrix based on spatial lead distribution. Finally, the three-dimensional feature matrix is processed through a capsule network with a hybrid module (CBAM-CapsNet) for training and prediction. Experimental results indicate that higher frequency bands have a more significant impact on emotion recognition, and the use of four-frequency band three-dimensional matrix can significantly enhance the accuracy of emotion recognition. CBAM-CapsNet achieved a binary classification accuracy of 95.42% and 95.52% for Arousal and Valence, respectively, on the DEAP dataset, a four-class accuracy of 95.00% for Arousal_Valence combined on the DEAP dataset, and a three-class accuracy of 93.19% on the SEED dataset. Compared to existing mainstream brain emotion recognition models based on deep learning, the accuracy is significantly improved.