LAUSR

Data augmentation is an effective way to increase the diversity of existing training datasets that result in improved generalization ability of convolutional neural networks (CNNs). The augmentation effect is usually global for the existing methods i.e., a single augmentation effect is applied to the whole image, thus limiting the diversity of local characteristics in augmented images. Moreover, the global augmentation effect does not support the most fundamental behavior of CNNs i.e., they focus more on local features (local texture, tiny noise etc.) than global shapes. We refer to this behavior as local bias property. In this paper, we propose a new data augmentation method, called Local Augment (LA), which highly alters the local bias property so that it can generate significantly diverse augmented images and offers the network with a better augmentation effect. First, we select few local patches in an image, then apply different types of augmentation strategies to each local patch. This augmentation process collapses the global structure of the object but creates locally diversified samples, which helps the network to learn the local bias property in a more generalized way. As a result, it increases the generalizability and the prediction accuracy of the network. To verify the effectiveness of the proposed method, we perform comprehensive experiments on image classification with benchmark datasets, where the proposed method outperforms the sate-of-the-art data augmentation techniques on ImageNet and STL10 and shows competitive performance on CIFAR100.