Figure 5: Left: 10-fold cross-validated CNN classification results for digit class ’6’ based on different original training data
sample sizes for digit ’6’ and augmented with various data generation strategies. Right: Zoom into left plot for class ’6’s
training data sample sizes 10 to 300 and corresponding classification accuracies ≥ 0.9. Note the scale of y-axes.
dures, respectively. The analysis of classification per-
formance of each scenario was carried out on a shared
real test sample of digit images. Results showed
that classical digit rendering contributes a lot to im-
provements already for very scarce training samples.
However, GAN refined versions of those data yet
still elevated performances significantly. Moreover,
it pays off investing into the rendering model, e.g. by
background modeling. GAN augmentation strategies
based only on available real training images do have a
positive impact on classification, but to a much lesser
extent and not necessarily reliably.
Finally, our results revealed the simple truth that
the more possibly realistic information is put into an
augmentation strategy the better will be the final clas-
sification outcome. While GANs are often reported
of yielding astonishing results, their performances
are only as good as the underlying data they operate
on. However, combined with image rendering, they
clearly showed the capability of overcoming signifi-
cant missing training data scenarios.
REFERENCES
Antoniou, A., Storkey, A., and Edwards, H. (2017). Data
augmentation generative adversarial networks. arXiv
preprint arXiv:1711.04340.
Arjovsky, M. and Bottou, L. (2017). Towards principled
methods for training generative adversarial networks.
arXiv preprint arXiv:1701.04862.
Arjovsky, M., Chintala, S., and Bottou, L. (2017). Wasser-
stein generative adversarial networks. In International
Conference on Machine Learning, pages 214–223.
Barratt, S. and Sharma, R. (2018). A note on the inception
score. arXiv preprint arXiv:1801.01973.
Glorot, X. and Bengio, Y. (2010). Understanding the diffi-
culty of training deep feedforward neural networks.
In Proceedings of the thirteenth international con-
ference on artificial intelligence and statistics, pages
249–256.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Ben-
gio, Y. (2014). Generative adversarial nets. In
Advances in neural information processing systems,
pages 2672–2680.
Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., and
Courville, A. C. (2017). Improved training of wasser-
stein gans. In Advances in Neural Information Pro-
cessing Systems, pages 5769–5779.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In Proceedings of
the IEEE conference on computer vision and pattern
recognition, pages 770–778.
Karras, T., Aila, T., Laine, S., and Lehtinen, J. (2017). Pro-
gressive growing of gans for improved quality, stabil-
ity, and variation. arXiv preprint arXiv:1710.10196.
Lucic, M., Kurach, K., Michalski, M., Gelly, S., and Bous-
quet, O. (2017). Are gans created equal? a large-scale
study. arXiv preprint arXiv:1711.10337.
MATLAB (2016). version 9.1.0 (R2016b). The MathWorks
Inc., Natick, Massachusetts.
More, A. (2016). Survey of resampling techniques for
improving classification performance in unbalanced
datasets. arXiv preprint arXiv:1608.06048.
Pinetz, T., Soukup, D., Huber-M
¨
ork, R., and Sablatnig, R.
Using a u-shaped neural network for minutiae extrac-
tion trained from refined, synthetic fingerprints.
Radford, A., Metz, L., and Chintala, S. (2015). Unsu-
pervised representation learning with deep convolu-
tional generative adversarial networks. arXiv preprint
arXiv:1511.06434.
Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:
Convolutional networks for biomedical image seg-
mentation. In International Conference on Medical
image computing and computer-assisted intervention,
pages 234–241. Springer.
Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V.,
Radford, A., and Chen, X. (2016). Improved tech-
ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods
22