Adversarial Learning of General Transformations for Data Augmentation

Data augmentation is essential for preventing overfitting in large convolutional neural networks with limited training data. Rather than applying predefined transformations, this work learns augmentation directly from training data using an encoder-decoder architecture combined with spatial transformer networks. The resulting transformed images maintain their original class while providing new, more complex samples for classifier training. Experimental results demonstrate superiority over previous generative augmentation approaches and comparable performance to traditional transformation methods.