Electrotactile stimulation may cause severe interference with the acquisition and analysis of surface electromyography (sEMG) signals, adversely affecting the control performance of sEMG-based prosthetic hands. Therefore, it is necessary to remove the interference from the electrical stimulation signal to enhance the accuracy of sEMG-based gesture recognition under electrotactile feedback. This paper proposes an electrical stimulation artifact removal method based on function interpolation. An sEMG gesture database was established, including three types of signals—original raw-tata, filtered-data and comblined-data. Experiment results show that the proposed method outperforms wavelet thresholding and adaptive filtering in terms of signal-to-noise ratio (15.52) and root mean square error (5.01). Vision Transformer (ViT) classification model was introduced for gesture recognition. The offline recognition accuracy of sEMG signals influenced by electrotactile feedback was increased from about 60% to about 90%. Online experiments were further conducted to validate its artifact de-noising performance.