Masking the Backgrounds to Produce Object-Focused Images and Comparing that Augmentation Method to Other Methods

Abstract

Image augmentation is a very powerful method to expand existing image datasets. This paper presents a novel method for creating a variation of existing images, called Object-Focused Image (OFI). This is when an image includes only the labeled object and everything else is made white. This paper elaborates on the OFI approach, explores its efficiency, and compares the validation accuracy of 780 notebooks. The presented testbed makes use of a subset of ImageNet Dataset (8,000 images of 14 classes) and incorporates all available models in Keras. These 26 models are tested before augmentation and after applying 9 different categories of augmentation methods. Each of these 260 notebooks is tested in 3 different scenarios: scenario A (ImageNet weights are not used and network layers are trainable), scenario B (ImageNet weights are used and network layers are trainable) and scenario C (ImageNet weights are used and network layers are not trainable). The experiments presented in this paper show that using OFI images along with the original images can be better than other augmentation methods in 16.4% of the cases. It was also shown that OFI method could help some models learn although they could not learn when other augmentation methods were applied. The conducted experiments also proved that the Kernel filters and the color space transformations are among the best data augmentation methods.