In recent years, gesture recognition based on surface electromyography (sEMG) signals has attracted the attention of researchers due to its non-invasive and high efficiency. However, the traditional gesture recognition methods still have limitations in the accuracy of recognizing complex gesture scenarios. In addition, most existing methods only model in the time or spatial dimension, and cannot fully utilize the frequency information of the signal. Although the common attention mechanism (such as SE module) can enhance the importance of channel features, it is limited to global average pooling and easily loses the spatial information within the channel. To address these issues, this paper proposes a hand gesture recognition framework named FE-ResNet50, which combines frequency-enhanced channel attention mechanism (FECAM) with ResNet50 backbone network. The network places the channel attention mechanism before the residual layer of ResNet50 network, while capturing spatial features of hand gesture signals through the deep network structure of ResNet50. By training on a self-built dataset, the network achieves an average classification accuracy of 92.77%. The validation on the NinaProDB1 public dataset shows that the network achieves an accuracy of 88.93%. Experimental results show that the network can effectively improve the performance of sEMG-based gesture recognition, has good robustness and generalization ability, and shows good performance and application potential in actual gesture recognition tasks.