In response to the need to improve the accuracy of gesture recognition for multi-channel surface electromyography(sEMG)sensors

a new gesture recognition framework is proposed

using continuous wavelet transform(CWT)to convert signals into time-frequency maps and designing the DIFT-Net classification network

integrating the Swin Transformer dual-branch structure

enhancing local feature extraction through the attention guidance module(AGM)

and achieving multi-scale global feature fusion through the attention interaction module(AIM).Experiments are verified based on the self-built datasets and Nanapro DB1 datasets

and DIFT-Net achieves accuracies of 98.36% and 82.64%

respectively

which is significantly superior to traditional methods.The results show that the framework can effectively integrate local and global features

significantly improve the gesture recognition performance of multi-channel sEMG sensors

and provide a new solution for human-computer interaction applications.