Current research on gesture recognition based on convolutional neural network (CNN) focuses on increasing the depth of the network, and pays less attention to improve the distribution of sample data which can brought the performance improvement. Aimed at these problems, a kind of electromyography feature matrix ( EFM) sample that quantifies the correlation of surface electromyography (sEMG) features is fed into the efficient channel attention (ECA) mechanism CNN, which is used to identify the 52 types gesture in NinaproDB1. Firstly, the time window is used to truncate the low-pass filtered sEMG and calculate various signal time domain features. Then, the cartesian product is used to combine and multiply different features. The EFM is obtained after normalizing the feature multiplication values. At the same time, ECA mechanism is introduced to make the network focus on the important deep features, thereby improving the effect of gesture classification. sEMG, EMG time-domain features and EFM are fed to the attention mechanism CNN respectively for gesture recognition. The recognition accuracy of EFM is the highest and reached 86. 39%, which is higher than the accuracy of gesture recognition research methods in recent years. The effectiveness of the proposed method is verified, and a feasible new scheme for accurate multi-category gesture classification is provided.