Onion Rotation: An Information-reserved Rotation Method to Augment Data in Image Recognition

摘要

With the amount of data dramatically increasing, models with countless parameters are naturally proposed to fit those massive data. As a basic form of data augmentation, rotation is widely used to generate new training samples. Effective as the standard rotation is, it commonly loses image information thus harmful to model performance sometimes. In this work, we propose a new image rotating technique which is called Onion Rotation. We consider the image as an "onion" which is composed of concentric circles and each circle is rotated separately. On the one hand, this rotation policy combines advantages of several standard data augmentation methods without information lost; on the other hand, Onion Rotation applies different augmentation policies to different pixels in the image. Therefore, our method creates great variations globally and locally, performing as a stronger regularization than standard augmentation methods. Moreover, Onion Rotation is easy and efficient to implement. We evaluate our method with ResNet and ResNet with the Squeeze and Excitation module on CIFAR-10, CIFAR-100 and SVHN datasets, comparing with several related data augmentation approaches. The experiment results show that our technique can fairly promote model performance solidly.