Computer Vision News - September 2020

Adversarial Data Augmentation... 7 image classification and segmentation . As you may remember from last month’s article, the main reason of data augmentation is important, especially in the medical imaging field, as there are often not enough data to train deep learning models, which require a vast amount of labelled datasets for training . Usually, geometric and intensity random transformations are used from the original images to perform accurately medical tasks (diagnosis, segmentation, classification). One issue with the traditional approach is that the modified (“augmented”) images have similar distributions to the originals. GANs have been previously used to improve the performance of such task-oriented deep learning models, as they can synthesise realistic and novel samples with good ability to generalise by exploiting the uncovered real image distribution . As a quick reminder, the basic idea of GANs is that given 2models, a generativemodel and a discriminative model , the task is to determine whether a given image appears natural or looks like it has been artificially created. The generator creates natural- looking images, similar to the original data distribution (minimax). The generator is trying to deceive the discriminator while the discriminator is trying to not get deceived by the generator . During model training, the alternating optimisation leads to improvement , good enough not to be able to distinguish fake from real images , as it is shown in Fig. 2. Figure 2. GANs architecture using as an example an input image of a violet flower.