Hadamard Walsh space based hybrid technique for image data augmentation

Abstract

Image data augmentation (IDA) is common when deep learning is used for image classification to address the issue of overfitting. Overfitting occurs when the datasets are small and the deep learning models have a huge capacity. Overfitting models have low training errors but high validation errors and result in poor generalization. Several methods have been researched in this context, but frequency domain-based methods are less explored. In this research, we have explored the Hadamard and Walsh space and developed two hybrid technique for IDA. The proposed techniques use a combination of Hadamard/Walsh transform and geometrical transformations. Empirical study is carried out using the VGG-16 model for image classification on the CIFAR-10 dataset and the results are compared with existing methods. The analysis of the results shows that the proposed techniques improve the evaluation parameters significantly. Further, analysis of training loss vs. validation loss shows that the proposed Hadamard-based hybrid methods have better generalization ability than the proposed Walsh-based hybrid method.