An evolutionary approach to automated class-specific data augmentation for image classification

Convolutional neural networks (CNNs) can achieve remarkable performance in many computer vision tasks (e.g. classification, detection and segmentation of images). However, the lack of labelled data can significantly hinder their generalization capabilities and limit the scope of their applications. Synthetic data augmentation (DA) is commonly used to address this issue, but uniformly applying global transformations can result in suboptimal performance when certain changes are more relevant to specific classes. The success of DA can be improved by adopting class-specific data transformations. However, this leads to an exponential increase in the number of combinations of image transformations. Finding an optimal combination is challenging due to a large number of possible transformations (e.g. some augmentation libraries offering up to sixty default transformations) and the training times of CNNs required to evaluate each combination. Here, we present an evolutionary approach using a genetic algorithm (GA) to search for an optimal combination of class-specific transformations subject to a feasible time constraint. Our study demonstrates a GA finding augmentation strategies that are significantly superior to those chosen randomly. We discuss and highlight the benefits of using class-specific data augmentation, how our evolutionary approach can automate the search for optimal DA strategies, and how it can be improved.

This version of the paper has been accepted for publication, after peer review and is subject to Springer Nature’s AM terms of use (https://www.springernature.com/gp/open-research/policies/accepted-ma...), but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at: http://dx.doi.org/10.1007/978-3-031-50320-7_12