A comparison of graph-based semi-supervised learning for data augmentation

Author(s)

de Oliveira, Willian Dihanster G. ; Penatti, Otavio A. B. ; Berton, Lilian ; IEEE

• Document type: Journal article
• Source: 2020 33RD SIBGRAPI CONFERENCE ON GRAPHICS, PATTERNS AND IMAGES (SIBGRAPI 2020); v. N/A, p. 8-pg., 2020-01-01.

Abstract

In supervised learning, the algorithm accuracy usually improves with the size of the labeled dataset used for training the classifier. However, in many real-life scenarios, obtaining enough labeled data is costly or even not possible. In many circumstances, Data Augmentation (DA) techniques are usually employed, generating more labeled data for training machine learning algorithms. The common DA techniques are applied to already labeled data, generating simple variations of this data. For example, for image classification, image samples are rotated, cropped, flipped or other operators to generate variations of input image samples, and keeping their original labels. Other options are using Neural Networks algorithms that create new synthetic data or to employ Semi-supervised Learning (SSL) that label existing unlabeled data. In this paper, we perform a comparison among graph-based semi-supervised learning (GSSL) algorithms to augment the labeled dataset. The main advantage of using GSSL is that we can increase the training set by adding nonannotated images to the training set, therefore, we can benefit from the huge amount of unlabeled data available. Experiments are performed on five datasets for recognition of handwritten digits and letters (MNIST and EMINIST), animals (Dogs vs Cats), clothes (MNIST-Fashion) and remote sensing images (Brazilian Coffee Scenes), in which we compare different possibilities for DA, including the GSSL, Generative Adversarial Networks (GANs) and traditional Image Transformations (IT) applied on input labeled data. We also evaluated the impact of such techniques on different convolutional neural networks (CNN). Results indicate that, although all DA techniques performed well, GSSL was more robust to different image properties, presenting less accuracy variation across datasets. (AU)

• FAPESP's process: 18/01722-3 - Semi-supervised learning via complex networks: network construction, selection and propagation of labels and applications
• Grantee: Lilian Berton
• Support Opportunities: Regular Research Grants