The Weakly Supervised Network of Hierarchical Attention Mechanism for Fine-Grained Classification

Fine-grained classification is challenging task to discriminate subtle and local differences from sub-categories. Many works improve the accuracy by relying heavily upon the use of the object or part annotations of images whose label are costly. In the paper, a weakly supervised network is proposed for fine-grained image classification without using expensive annotations. Firstly, it learns object detector by hierarchical attention mechanism automatically and localizes the objects or its parts to extract salient feature. Then, based on the theory of prototypical networks, it learns a metric loss function by computing distances to prototype representations of each class which is unsupervised clustering methods. Both are jointed to remove useless information or noise patches and retain discriminative features. Finally, we apply the proposed method to complete the classification of disaster-scene images. Compared with other methods, experimental results demonstrate our method is generalized and robust to different datasets.