The gesture recognition using electromyography ( EMG) has advantages of selective detail information and strong antiinterference ability. However, the adaptability and recognition accuracy of the existing methods are insufficient. By adding a long-term and short-term memory network layer on the basis of the convolutional neural network, a gesture recognition model is formulated. In this way, it can capture the EMG timing characteristics of the gesture, and the phenomenon of overfitting is reduced to a certain degree. The rich time-frequency domain information of EMG is utilized to extract the wavelet packet feature image of EMG. In addition, the input data of the recognition model are used with the EMG image to expand the category information of the EMG signal. Meanwhile, the attention mechanism is introduced between the time memory network processing layer and the convolutional neural network layer. Then, the model can indirectly increase the weights of the key gesture EMG channels. Compared with the method of ordinary convolutional neural network model using single EMG image, experimental results show that the recognition accuracy rate of the processing methods of EMG two feature inputs is improved by 4. 25% .