SAPDA: Significant Areas Preserved Data Augmentation

Data Augmentation is an essential technology for improving the performance of deep learning models. However, the semantic information change in current data augmentation methods may impair the model performance, especially in randomly erasing-based data augmentation. We focus on exploiting a Significant Areas Preserved Data Augmentation (SAPDA) method to mitigate this issue, where the most informative areas are preserved. Moreover, the significant areas (SA) are derived from perceptions of current models during the model training process. Therefore, the areas preserved by SAPDA are different for each stage of training. Inspired by hard sample mining, first, we define the SA of images by employing a Class Activation Map (CAM). The SA consists of Discriminative Area (DA), Misclassified Area (MA), and Uncertain Area (UA). Further, the SA is preserved during data augmentation, which could keep the semantic information unchanged and make the model focus on difficult and uncertain areas. Extensive experiments are conducted on CIFAR-10, CIFAR-100, and ImageNet, which show that our proposed SAPDA could further improve the model performance in combination with existing SOTA data augmentation methods. Also, the SAPDA could enhance the model’s ability in semi-supervised learning and noisy label learning.