Research On Convolutional Neural Network Based On Rotation Invariant Feature

Convolutional neural network(CNN)has achieved great success in feature extraction of original input data such as images.Although traditional convolutional neural networks are invariant to translation on input,they are not invariant to other transformations,including rotation and flipping.Generally,global rotation invariance is usually sought through data augmentation,but it is very difficult for patch-wise to have rotation invariance.Aiming at the difficulty that CNN does not have rotation invariance,this paper proposes a convolution neural network topology based on image moments,which combines image moments and convolution networks.This paper proves that it can effectively deal with prior knowledge about rotation changes in data and requires fewer model parameters and training data than usual data augmentation methods.On the other hand,a new form of convolution kernel based on image moment decomposition is implemented,which divides the traditional convolution layer into two steps.In the first step,image moment coefficient is used to filter out the information which is least sensitive to rotation to form a feature map.The second step is to connect the convolution layer with 1*1 convolution core to extract higher dimensional information and use it for back propagation training.This aggregation operation makes the extracted information insensitive to rotation direction and improves the ability of rotation generalization.At the same time,this paper proposes a fast method to calculate image moment features,which uses the conversion of time-domain convolution and frequency-domain product for fast calculation,so as to improve the calculation efficiency and reduce the calculation cost.This paper tests the rotation invariance performance on MNIST and MNIST-rot data sets,EMNIST and EMNIST-rot data sets,and compares with the two baseline network structures designed and the mainstream methods.It has excellent performance in predicting the rotation version of the original image on two data sets,especially when training on the model based on the original MNIST and testing on the MNIST-rot.We prove that the convolution neural network based on image moments proposed in this paper finds the most ideal rotation orientation information in the input image for training.With fewer parameters,it can achieve better accuracy,and the network structure is simpler,which is more conducive to utilization.