Image Periodization for Convolutional Neural Networks

In last few years, convolutional neural networks (CNNs) have achieved great success on many image classification tasks, which shows the effectiveness of deep learning methods. The training process of CNN is usually based on a large number of training data, and the CNN can learn local image translation invariance property by using the convolution operation and pooling operation, which is very important for image classification. However, the performance of CNN is limited when the location of the key object varies greatly in the image or the number of data is insufficient. Addressing this problem, in this paper, we propose a novel method named image periodization for CNN on image classification tasks. We extend the original image periodically and resample it to generate new images, while we design a circular convolutional layer to replace the original convolutional layer. Our method can be used as a data augmentation method, and it can provide complete translation invariance property for CNNs. Our method can be easily plugged into common CNNs, and the experiment results show consistent improvement on different CNN-based models.