A Model-Agnostic Derivative of Cutout for Image Data Augmentation

Abstract

Data augmentation is a strong tool for enhancing the performance of deep learning models using different techniques to increase both the quantity and diversity of training data. Cutout was previously proposed, in the context of image classification, as a simple regularization technique that increases Convolutional Neural Networks' robustness and performance by masking part of the input image. Cutout can also be applied along with other data augmentation or regularization techniques, further improving their effectiveness. We extend the main idea of Cutout, where instead of randomly masking images as Cutout does, we use a model to identify which part of the image affects the model's decision the most. Similar to Cutout, our new approach can be used with other regularization techniques for better performance. Our experiments using the CIFAR-10 and CIFAR-100 datasets with the same ResNet18 and WideResNet models and parameters used in the original Cutout paper show that we can get slightly higher accuracy but at a higher cost than Cutout. These results suggest that randomness is an effective driving factor in Cutout, making it accurate enough and time-efficient. Nonetheless, our approach can be helpful when accuracy is the main concern and the extra time cost is acceptable for the aimed application.