In classification tasks, the robustness against various image transformations remains a crucial property of CNN models. When acquired using the data augmentation it comes at the price of a considerable increase in training time and the risk of overfitting. Consequently, researching other ways to endow CNNs with invariance to various transformations is an intensive field of study.

This paper presents a new reduced, rotation-invariant, classification model composed of two parts: a feature representation mapping and a classifier. We provide an insight into the principle and we show that the proposed model is trainable. The model we obtain is smaller and has angular prediction capabilities.

We illustrate the results on the MNIST-rot and CIFAR-10 datasets. We achieve the state-of-the-art classification score on MNIST-rot, and improve by 20% the state of the art score on rotated CIFAR-10. In all cases, we can predict the rotation angle.

在分类任务中，针对各种图像变换的鲁棒性仍然是CNN模型的关键属性。当使用数据增强来获取数据时，将以大量增加训练时间和过度拟合的风险为代价。因此，研究赋予CNN各种变换不变性的其他方法是一个密集的研究领域。

本文提出了一种新的归约旋转不变分类模型，该模型由两部分组成：特征表示映射和分类器。我们提供了对该原理的深入了解，并表明所提出的模型是可训练的。我们获得的模型较小，并且具有角度预测功能。

我们说明了MNIST-rot和CIFAR-10数据集的结果。我们在MNIST-rot上达到了最新的分类评分，在旋转的CIFAR-10上又提高了20％的最新评分。在所有情况下，我们都可以预测旋转角度。