Attentive CutMix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification

Convolutional neural networks (CNN) are capable of learning robust\nrepresentation with different regularization methods and activations as\nconvolutional layers are spatially correlated. Based on this property, a large\nvariety of regional dropout strategies have been proposed, such as Cutout,\nDropBlock, CutMix, etc. These methods aim to promote the network to generalize\nbetter by partially occluding the discriminative parts of objects. However, all\nof them perform this operation randomly, without capturing the most important\nregion(s) within an object. In this paper, we propose Attentive CutMix, a\nnaturally enhanced augmentation strategy based on CutMix. In each training\niteration, we choose the most descriptive regions based on the intermediate\nattention maps from a feature extractor, which enables searching for the most\ndiscriminative parts in an image. Our proposed method is simple yet effective,\neasy to implement and can boost the baseline significantly. Extensive\nexperiments on CIFAR-10/100, ImageNet datasets with various CNN architectures\n(in a unified setting) demonstrate the effectiveness of our proposed method,\nwhich consistently outperforms the baseline CutMix and other methods by a\nsignificant margin.\n