Abstract: Gesture recognition is a hot topic in the fields of human-computer interaction, prosthesis control, rehabilitation, and so on. To satisfy the requirement of real-time and accuracy on gesture recognition, this paper makes the analysis on LeNet-5 and proposes the LeNet-A network suitable for acceleration signal. In the proposed LeNet-A, the low-cost acceleration signal is taken as the input of network, which are six layers, including two convolution layers, two pooling layers, one fully connected layer and one output layer. Especially, by considering the specific complexity of gesture recognition based on acceleration signal, we introduce the Dropout layer and change the size of convolution kernel and the number of convolution kernel. Besides, the activation function is replaced by Relu function and the classifier is replaced by Softmax classifier. This paper chooses Ninapro database as the dataset, which is an available public resource for research on gesture recognition and has up to more than 50 gestures including intact subject and the amputated. Before making classification, these signals in this dataset are dealt with by low-pass filtering, down-sampling, data balance and training/testing set segmentation. The simulation results show that the proposed network has great advantage in the gesture recognition both in the intact subject and the amputated and attain the average accuracy of 90.37% and 79.99%, respectively, which increase by 12% and 31%, compared with the best result in the exiting literature. The test on one subject with and without rest gesture shows that the accuracy without rest gesture is 93.60%, increasing by 3.1% compared with the accuracy with rest gesture. It is observed that the rest gesture has a relatively large influence on the whole gesture classification accuracy. The latter experiment also illustrates that this proposed network not only improves the gesture recognition accuracy distinctly with characteristics of good real-time, but also has quite strong robustness to noise. These further verify that the proposed method in this paper is quite significant for gesture recognition.