Abstract: Existing EEG-based emotion recognition methods still suffer from limited local feature representation capability and inefficient feature extraction, which hinder further improvements in recognition performance. To address these limitations, an emotion recognition model integrating a multi-scale feature pyramid and attention mechanism, termed the Multi-Scale Pyramid Attention Transformer (MSPAT), is proposed. Firstly, a multi-scale convolution-based feature pyramid module is designed to extract multi-scale spatial and frequency-domain features from EEG topological maps. Subsequently, a multi-scale attention aggregation module is introduced to enhance emotion-rela-ted discriminative features in key channels and spatial regions. Finally, a temporal Transformer module is employed to model the temporal dynamics of emotional variations. To evaluate the effectiveness and stability of the proposed model, comparative experiments were conducted on the DEAP and SEED datasets against representative mainstream baseline models, including 4D-CRNN, STFCGAT, and Conformer, together with confusion matrix analysis and ablation studies. Experimental results demonstrate that MSPAT achieves classification accuracy of 97.61% and 97.12% on the Valence and Arousal dimensions of the DEAP dataset, respectively, with improvements of 0.8%-7.8% and 0.5%-6.5% over all baseline models in the two dimensions; it obtains an classification accuracy of 96.17% on the SEED dataset, representing an improvement of 0.9%-12.4% compared with all baseline models. Meanwhile, the results of ablation studies further verify the effectiveness of the proposed multi-scale feature pyramid module and multi-scale attention module in enhancing local feature representation capability and improving feature extraction efficiency. The results indicate that MSPAT can fully mine the deep correlation information of EEG signals, effectively alleviate the problems of insufficient feature expression and incomplete information mining, and exhibit exce-llent recognition stability and cross-dataset generalization performance.