Rotation-equivariant convolutional neural network ensembles in image processing

ABSTRACT

For the present engineering of neural networks, rotation invariant is hard to be obtained. Rotation symmetry is an important characteristic in our physical world. In image recognition, using rotated images would largely decrease the performance of neural networks. This situation seriously hindered the application of neural networks in the real-world, such as human tracking, self-driving cars, and intelligent surveillance. In this paper, we would like to present a rotation-equivariant design of convolutional neural network ensembles to counteract the problem of rotated image processing task. This convolutional neural network ensembles combine multiple convolutional neural networks trained by different ranges of rotation angles respectively. In our proposed theory, the model lowers the training difficulty by learning with smaller separations of random rotation angles instead of a huge one. Experiments are reported in this paper. The convolutional neural network ensembles could reach 96.35% on rotated MNIST datasets, 84.9% on rotated Fashion-MNIST datasets, and 91.35% on rotated KMNIST datasets. These results are comparable to current state-of-the-art performance.

References

1. Xiuyuan Cheng, Qiang Qiu, Robert Calderbank, and Guillermo Sapiro. 2018. RotDCF: Decomposition of Convolutional Filters for Rotation-Equivariant Deep Networks. arXiv preprint arXiv:1805.06846 (2018).
2. Taco Cohen and Max Welling. 2016. Group equivariant convolutional networks. In International conference on machine learning. 2990--2999.
3. Beranger Dumont, Simona Maggio, and Pablo Montalvo. 2018. Robustness of rotation-equivariant networks to adversarial perturbations. arXiv preprint arXiv:1802.06627 (2018).
4. James D Foley, Foley Dan Van, Andries Van Dam, Steven K Feiner, John F Hughes, J HUGHES, and EDWARD ANGEL. 1996. Computer graphics: principles and practice. Vol. 12110. Addison-Wesley Professional.
5. Liyao Gao. 2019. Learning with Collaborative Neural Network Group by Reflection. arXiv preprint arXiv:1901.02433 (2019).
6. Robert A Jacobs, Michael I Jordan, Steven J Nowlan, Geoffrey E Hinton, et al. 1991. Adaptive mixtures of local experts. Neural computation 3, 1 (1991), 79--87.
7. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems. 1097--1105.
8. Yann LeCun, Yoshua Bengio, et al. 1995. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks 3361, 10 (1995), 1995.
9. Junying Li, Zichen Yang, Haifeng Liu, and Deng Cai. 2018. Deep rotation equivariant network. Neurocomputing 290 (2018), 26--33.
10. Diego Marcos, Michele Volpi, Nikos Komodakis, and Devis Tuia. 2017. Rotation equivariant vector field networks. In Proceedings of the IEEE International Conference on Computer Vision. 5048--5057.
11. Andrew Michael Saxe, Yamini Bansal, Joel Dapello, Madhu Advani, Artemy Kolchinsky, Brendan Daniel Tracey, and David Daniel Cox. 2018. On the information bottleneck theory of deep learning. (2018).
12. Maurice Weiler, Fred A Hamprecht, and Martin Storath. 2018. Learning steerable filters for rotation equivariant CNNs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 849--858.
13. Zhi-Hua Zhou, Jianxin Wu, and Wei Tang. 2002. Ensembling neural networks: many could be better than all. Artificial intelligence 137, 1--2 (2002), 239--263.

Index Terms

• Computing methodologies

  • Machine learning

    • Machine learning algorithms

      • Ensemble methods

    • Machine learning approaches

      • Neural networks