이미지 패치 연결 방식을 통한 데이터 증강 기술의 경향성에 관한 연구

Abstract

The convolutional neural networks (CNNs) for image classification tasks are getting larger and deeper. While CNNs have good representation power using numerous parameters, overfitting may occur if the size of the training dataset is insufficient. Therefore, various data augmentation techinques such as rotation, flipping, and cropping have been used to prevent overfitting. Recently, novel data augmentation techniques such as CutMix and RICAP achieved state-of-the-art performance by combining multiple images into one image. Even though he performance is known to depend on the number of images for composition, such dependency has not been sufficiently studied. In this study, we discuss tendencies of the data augmentation technique using the image patching method. We experimented validation error rates for various image patches and patch counts. Experiments are conducted using two convolution-based networks: ResNet, WideResNet on the CIFAR-10 and mini-ImageNet datasets.