Flipping Data Augmentation of Convolutional Neural Networks Using Discrete Cosine Transforms
Izumi Ito · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
Convolutional neural networks (CNNs) are widely used in many areas. The problem now is to collect large numbers of labeled images in order to improve network performance. Data augmentation is to increase the number of images for training, where images are artificially generated by transformation, such as rotation, translation, scaling, and flipping. In this paper, we focus on flipping data augmentation and present a novel algorithm of convolution that involves flipping data augmentation in CNNs. Without generating flipped images beforehand, we can obtain information of flipped images in computation on convolutional layers using discrete cosine transforms. The proposed algorithm on a simple CNN is demonstrated and the efficacy of the proposed algorithm is testified.