Target Cropping: A New Data Augmentation Method of Fine-Grained Image Classification

Abstract

In this paper, we propose a novel data augmentation method of fine-grained image classification named target cropping. Previous work has demonstrated the effectiveness of data augmentation through simple technique, such as random cropping, image rotating and image flipping. But for fine-grained classification, due to its inter-class similarity and intra-class differences, traditional random cropping does not pay much attention to the discriminative regions and even may crop out the regions that have a critical impact on the classification results. To solve this problem, we propose target cropping which uses class activation maps to locate discriminative region. Compared with random cropping, our method significantly improves classification accuracy for all the tested datasets. For example, classification accuracy is improved from 71.1% to 73.9% for CUB200-2011 dataset with VGG-16 and from 77.2% to 79.0% in the FGVC-Aircraft dataset. It is a significant improvement in fine-grained image classification field.