Rotation Equivariance Of Deep Convolutinal Neural Network

Deep Convolutional Neural Network has become the most popular model in computer vison in recent years.One of the important reasons is the translation equivalence of Convolutional Neural Networks.Many researchers try to propose rotation equivariant framework to improve Convolutional Neural Network.Most of equivariance works focus on learning shallow rotation equivariant feature rather than deep rotation equivariant feature.However,the state-of-the-art network always have more than 100 layers.To learn deep rotation equivariant feature,we propose Deep Rotation Equivariant Network consisting of cycle layers,isotonic layers and decycle layers.Based on specific weight rotation share mechanism,our three proposed layers could learn rotation covariant feature,keep deep feature rotation covariant and learn rotation equivariant feature from rotation covariant feature separately.We evaluate DRENs on Rotated MNIST and CIFAR-10 datasets and demonstrate that it can improve the performance of state-of-the-art architectures.