Surface electromyogram (sEMG) armband with electrodes mounted around the user's forearm is one of the most ergonomic wearable EMG devices. Definitely, the signal distribution varies greatly across different wearing positions of the armband based on physiological characters of the sEMG, which may compromise the performance and even make it inapplicable to use a recognition model built in one position. Thus, the electrode shift problem hinders wide applications of the sEMG armband, especially during its long-term and repetitive usage. High-density surface electromyogram signals recorded from a 2-dimensional electrode array carry rich spatial information and can be viewed as a muscular activation image. In the proposed method, a self-designed convolutional neural network (CNN) combined with data augmentation operation is used to learn muscular activity patterns at an original/baseline position of the electrode array. The method cropped the sEMG feature images for data augmentation. The CNN could learn the invariance of electrode shift from the cropped images. The performance of the proposed method was evaluated with data recorded by an armband-like sensor apparatus with five electrode arrays embedded in its inner side, which was placed over forearm of 1 subject performing 8 gestural tasks. The proposed method was able to maintain high task classification accuracies around 95% and outperformed two traditional methods (accuracies ranging from 50% to 75%) under conditions with 4 designated shifts.