Image data augmentation

More advanced augmentation techniques use neural and convolutional networks to generate new data. Therefore, the SMOTE approach is based on the synthetic generation of new features using the “nearest neighbor” method. Concerning images, the use of an encoder/decoder allows you to extract features from images, generate new ones, and transform them into an image, thus obtaining a new dataset element. Based on the adversarial approach used in GAN networks, new images can be created end-to-end. In addition, the use of the adversarial principle makes it possible to generate special noise filters leading to incorrect classification of images, which certainly makes it possible to increase reliability against attacks, and the accuracy of solutions created in conditions of provocations.