A Saliency-based Weakly-supervised Network for Fine-Grained Image Categorization

Fine-grained image categorization is to recognize subcategories from each other. In actual photographs, there is often large intra-class variation and small inter-class variation, which make fine-grained image categorization a challenging task. Existing bilinear-pooling-based methods are less interpretable, and visual-attention-based methods often suffer from limited number of salient part. In this paper, we propose a saliency-based weakly-supervised network, a simple approach that can effectively preserve all the salient details of the original image. Based on the saliency sampler backbone, we apply feature pyramid network structure to combine the saliency information with high-level features and use KL-divergence for knowledge distillation. Experiments on Caltech-UCSD Birds dataset show that components of our network are efficient and reliable.