Fine gesture recognition based on scalp EEG signal is an important technical means for motor function rehabilitation of stroke patients.In this paper, a self-attention-based Transformer model recognition method is improved by designing the actual gesture and taking the long range EEG signals in the motor functional area during finger fine movement as the data processing object. The Transformer model based on self-attention module and its variant top-K sparse Transformer model are constructed from time dimension and space dimension respectively, and a hybrid artifact removal method based on MEMD-CCA is designed combining the characteristics of EEG data structure. The improved Transformer model achieves excellent classification recognition results.