Paper acceptance at the IEEE International Joint Conference on Neural Networks (IJCNN).

Simple image rotations significantly reduce the accuracy of deep neural networks. Moreover, training with all possible rotations increases the data set, which also increases the training duration. In this work, we address trainable rotation invariant convolutions as well as the construction of nets, since fully connected layers can only be rotation invariant with a one-dimensional input. On the one hand, we show that our approach is rotationally invariant for different models and on different public data sets. We also discuss the influence of purely rotational invariant features on accuracy. The rotationally adaptive convolution models presented in this work are more computationally intensive than normal convolution models. Therefore, we also present a depth wise separable approach with radial convolution.