Home // eKNOW 2020, The Twelfth International Conference on Information, Process, and Knowledge Management // View article

Efficient Parameters for Rotation Processing of Data Augmentation

Authors:
Naoki Nakamura
Duke Maeda
Kenta Morita
Naoki Morita

Keywords: Convolutional neural network; Training data augmentation; Rotation angle; Augmentation rate.

Abstract:
Deep learning typically requires a large amount of training data. However, in some cases, it is not possible to prepare enough data to achieve the desired recognition accuracy. A number of approaches to training models with a limited amount of data are available, such as data augmentation and fine-tuning. In the present study, we focus on rotation processing, which has the capacity to augment image data more easily than other augmentation methods. With this method, for example, we could produce 360 images from a single image by rotating the image a full 360° by 1° increments. However, if the rotation angle is not chosen appropriately, essential features of the rotated object may be lost. No clear standards have been previously determined for setting appropriate rotation parameters. This study presents an approach to efficient rotary processing for cases in which the key features of the object either does or does not distort, depending on the angle of rotation. The approach should make it easier for general users to set proper parameters when using rotation processing for data augmentation.

Pages: 62 to 65

Copyright: Copyright (c) IARIA, 2020

Publication date: March 22, 2020

Published in: conference

ISSN: 2308-4375

ISBN: 978-1-61208-765-8

Location: Valencia, Spain

Dates: from November 21, 2020 to November 25, 2020