Introduction: Convolutional neural networks are widely used in gesture recognition research, which employs surface electromyography. However, when processing surface electromyography data, current deep learning models still face challenges, such as insufficient effective feature extraction, poor performance in multi-gesture recognition, and low accuracy in recognizing sparse surface electromyography.Methods: To address these issues, this study proposed a multi-stream adaptive convolutional neural networks with residual modules (MSACNN-RM) for surface electromyography gesture recognition, which integrates multiple streams of convolutional neural networks, adaptive convolutional neural networks, and residual modules to enhance the model’s feature extraction and learning capabilities. This improves the model’s ability to extract and understand complex data patterns.Results: The experimental results demonstrated that the model achieved recognition accuracies of 98.24%, 93.52%, and 92.27% respectively on the Ninapro DB1, Ninapro DB2, and Ninapro DB4 datasets. Compared with other deep learning models, MSACNN-RM achieves higher accuracy compared to existing models.Discussion: The proposed model explores features of sparse sEMG signals by leveraging multi-stream convolution, the combination of adaptive convolution modules and ResNet blocks enhances the model’s ability of extracting crucial gesture features. In the future, in order to deal with differences in sEMG signals caused by variations among individuals, a universal multi-gesture recognition algorithm should be developed. Meanwhile, the model should focus on optimizing and streamlining the network to reduce computational load.