A well-known type of data augmentation is image augmentation, and it is the technique of expanding the available training datasets that generates more training images after a series of random changes to the training images. An important goal of image augmentation is to produce a modified and transformed version of the image in the datasets that belong to the same class as the original image. Traditional image augmentation involves simple transformations such as horizontal flipping, color space augmentations, and random cropping. Currently, Image data augmentation is categorized into two approaches, namely basic image manipulations and deep learning approaches.
Techniques of Image augmentation based on basic image manipulation are geometric transformations, flipping, cropping, rotation, noise injection, translation, color space transformations, kernel filters, mixing images, and random erasing. As an advantage, the performance of deep learning neural networks improves with data availability and possesses the capability to fit the model.
Data augmentation-based deep learning approaches are feature space augmentation, adversarial training, GAN-based augmentation, and neural style transfer. The other technique for image data augmentation is meta-learning data augmentation which involves neural augmentation, smart augmentation, and auto augmentation. Recent and more useful application domains of image augmentation are medical image analysis and biomedical applications.