Our investigation aims to answer the research question is it possible to train deep network models that can be re-used to classify a new coming dataset of fingerprint images without re-training the new deep network model? For this purpose, we collect real datasets of fingerprint images from students at the Can Tho University. After that, we propose to train recent deep networks, such as VGG, ResNet50, Inception-v3, Xception, on the training dataset with 9,236 fingerprint images of 441 students, to create deep network models. And then, we re-use these resulting deep network models as the feature extraction and only fine-tune the last layer in deep network models for the new fingerprint image datasets. The empirical test results on three real fingerprint image datasets (FP-235, FP-389, FP-559) show that deep network models achieve at least the accuracy of 96.72% on the testsets. Typically, the ResNet50 models give classification accuracy of 99.00%, 98.33%, 98.05% on FP-235, FP-389 and FP-559, respectively.