A Unified Modular Framework with Deep Graph Convolutional Networks for Multi-label Image Recognition

With the rapid development of handheld photographic devices, a large number of unlabeled images have been uploaded to the Internet. In order to retrieve these images, image recognition techniques have become particularly important. As there is often more than one object in a picture, multi-label image annotation techniques are of practical interest. To enhance its performance by fully exploiting the interrelationships between labels, we propose a unified modular framework with deep graph convolutional networks (MDGCN). It consists of two modules for extracting image features and label semantic respectively, after which the features are fused to obtain the final recognition results. With classical multi-label soft-margin loss, our model can be trained in an endto-end schema. It is important to note that a deep graph convolutional network is used in our framework to learn semantic associations. Moreover, a special normalization method is employed to strengthen its own connection and avoid features from disappearing in the deep graph network propagation. The results of experiments on two multi-label image classification benchmark datasets show that our framework has advanced performance compared to the state-of-the-art methods.