Performance Analysis of DCT Based Latent Space Image Data Augmentation Technique

Image data augmentation is the process of enlarging the dataset by creating synthetic images using the existing images to address the problem of scarcity of data and poor generalization. The existing methods for image data augmentation are mainly domain specific and focus on transformations in spatial domain. In this paper, we have developed a domain independent image data augmentation technique based on frequency domain transformations. In this research, we have utilized discrete cosine transform (DCT)-based latent space and proposed two techniques. In the first technique, we apply DCT on each of the color channel of the image and synthesize new image by considering only high-energy latent coefficients. The second technique is a hybrid technique where standard transformations are applied on the images synthesized from the DCT latent space. The experiments carried out on the Cifar10 dataset using VGG16 revealed that the proposed hybrid technique is more effective than the technique where the synthetic images are created solely from the DCT-based latent space. The analysis of training loss and validation loss shows significant improvement in validation loss, thus improving generalization of the VGG 16 model.