Abstract: In the process of gesture recognition using EMG signals, affected by behavioral habits, muscle tissues, and different wearing styles, there were often difficulties such as long gesture training time, large data volume, low recognition accuracy, and poor real-time performance. For this reason, this paper used the migration learning theory to improve the LSTM algorithm model, selected different states of gestures to form the source task, took the trained LSTM network as the source network model, and completed the gesture recognition comparison experiments under the two migration strategies and non-migration strategies for the six types of gestures set. The results show that the recognition effect of the migration learning strategy using the pre-training approach is better than that of the migration strategy using the fixed-value approach. When the LSTM gesture recognition algorithm is improved by using the pre-training migration strategy, the average training time is only one-sixteenth of the training time required to use the LSTM recognition algorithm, and when each action is repeated only 20 times, the accuracy rate can reach 80. 2%, which is 22% higher on average than that of using only the LSTM gesture recognition method in the source network. Therefore, the gesture recognition accuracy can be improved while reducing the training amount by using the migration learning method.