MIDCN: A Multiple Instance Deep Convolutional Network for Image Classification

摘要

For the image classification task, usually, the image collected in the wild contains multiple objects instead of a single dominant one. Besides, the image label is not explicitly associated with the object region, i.e., it is weakly annotated. In this paper, we propose a novel deep convolutional network for image classification under a weakly supervised condition. The proposed method, namely MIDCN, formulate the problem into Multiple Instance Learning (MIL), where each image is a bag which contains multiple instances (objects). Different with previous deep MfL methods which predict the label of each bag (i.e., image) by simply performing pooling/voting strategy over their instance (i.e., region) predictions, MIDCN directly predicts the label of a bag via bag features learned by measuring the similarities between instance features and a set of learned informative prototypes. Specifically, the prototypes are obtained by a newly proposed Global Contrast Pooling (GCP) layer which leverages instances not only coming from the current bag but also the other bags. Thus the learned bag features also contain global information of all the training bags, which is more robust and noise free. We did extensive experiments on two real-world image datasets, including both natural image dataset (PASCAL VOC 07) and pathological lung cancer image dataset, and show the results of the proposed MIDCN consistently outperforms the state-of-the-art methods.