Currently, in the field of motor imagery decoding, research has focused on two approaches, subject-dependent and subject-independent decoding. However, these two decoding approaches have major limitations in the practical use of brain-computer interface (BCI) systems. Both subject-dependent and subject-independent decoding rely on the same center dataset, and when the decoding model is applied to datasets from other centers, the performance will be significantly degraded, which cannot satisfy the demand for cross-center use of BCI systems. To improve the cross-database decoding performance of motor imagery Electroencephalogram (EEG), a sparse selection model based on Fisher criterion regularization is proposed based on the methodological framework of domain generalization. Based on the least absolute shrinkage and selection operator (LASSO) model, a Fisher criterion regularization term is introduced to explicitly model feature separability during the feature selection process. This helps to improve the representation learning ability for domain generalization, thus enhancing the generalization performance of the classification model on different datasets. The effectiveness of the proposed method is validated using two publicly available motor imagery EEG datasets and two feature extraction methods, filter bank common spatial pattern (FBCSP) and multiple time frequency common spatial pattern (MTFCSP). Further validation through the utilization of self-collected data also confirmed the effectiveness of the proposed method in practical applications. Compared with existing methods, the proposed method achieved the highest average classification accuracy of 67.26%. The experimental results show that the proposed method has better generalization ability, higher feature separability, and better robustness in motion imagery cross-database decoding. The proposed method is expected to facilitate cross-center use of BCI systems and improve generalizability.