A Comparative Study for Natural Image Augmentation Methods

Abstract

This paper deals with natural image augmentation methods, i.e., those whose  results are close to the natural environmental effects that machine learning  models may encounter in industrial applications: from weather conditions; from  the operation or malfunction of environmental perception devices (e.g.  cameras); and so on. Structuring and analyzing natural image augmentation methods can bring significant practical relevance as a guide for researchers of  machine learning methods, as well as those developing machine learning models directly for industry. The currently known methods of natural image  augmentation in this paper have been divided into the following groups: those based on adding weather artifacts, camera artifacts or substituting the  background for the main object in the image. The paper also considers existing  software libraries for image augmentation: Albumentations, which has  integration with PyTorch; ImgAug, which provides a wide range of  possible natural augmentations; Augmentor, which provides the ability to add a variety of camera distortion effects to images. The CIFAR-10 dataset was selected for an experimental study of natural image augmentation techniques; augmentations were applied one at a time with a 50% probability each.