Data Augmentation using Evolutionary Image Processing

Abstract

In the machine learning community, data augmentation techniques have been widely used to make deep neural networks invariant to object transition. However, less attention has been paid to data augmentation in traditional classification methods. In this paper, we take a closer look at traditional classification methods and introduce a new data augmentation technique based on the concept of image transformation. Starting with a few existing examples, we add noise and generate new data points to reduce sparseness in a given feature space. Then, we generate images corresponding to the new data points, although this is usually an ill-posed problem. Herein, the novelty is in constructing an image transformation tree and generating new data from a small number of instances. This allows us to reduce sparseness in the feature space and build more robust classifiers. We evaluate our method on the Caltech-101 dataset to verify its potential. In the context of the situation where the amount of training data is limited, we demonstrate that the support vector machine-based classifiers trained with an augmented dataset using our method outperform classifiers trained with the original dataset in most cases.