arXiv (Cornell University) · 2019 · 158 citations · 0 references
ComputingComputer ScienceComputer VisionMachine LearningVisual AttentionAttention NetworksDeep LearningObject DetectionObject RecognitionOverfittingScene UnderstandingAttentionClassification AccuracyData AugmentationFine-grained Image ClassificationFine-grained Visual RecognitionImage CroppingWeakly Supervised Image ClassificationWeakly Supervised Learning
Data augmentation is usually adopted to increase the amount of training data,\nprevent overfitting and improve the performance of deep models. However, in\npractice, random data augmentation, such as random image cropping, is\nlow-efficiency and might introduce many uncontrolled background noises. In this\npaper, we propose Weakly Supervised Data Augmentation Network (WS-DAN) to\nexplore the potential of data augmentation. Specifically, for each training\nimage, we first generate attention maps to represent the object's\ndiscriminative parts by weakly supervised learning. Next, we augment the image\nguided by these attention maps, including attention cropping and attention\ndropping. The proposed WS-DAN improves the classification accuracy in two\nfolds. In the first stage, images can be seen better since more discriminative\nparts' features will be extracted. In the second stage, attention regions\nprovide accurate location of object, which ensures our model to look at the\nobject closer and further improve the performance. Comprehensive experiments in\ncommon fine-grained visual classification datasets show that our WS-DAN\nsurpasses the state-of-the-art methods, which demonstrates its effectiveness.\n