Label-aware graph representation learning for multi-label image classification

Multi-label image classification (MLIC) is a quintessential but challenging issue in the field of Computer Vision. Since the label co-occurrence is a crucial component of MLIC, previous existing approaches resort to the label co-occurrence for either modeling label correlations or modeling visual feature relationships. However, these methods ignore either the feature interaction or the label characteristics in MLIC. In this paper, we propose a label-aware graph representation learning (LGR) for MLIC that can explore the label interaction via a graph neural network built on the label co-occurrence and mine the feature correlations via another graph neural network also based on the label co-occurrence. Moreover, to decouple semantic visual features, current approaches resort to the word embedding guided semantic decoupling methods. However, the word embedding cannot clearly represent the label semantic information of MLIC. Hence, we reconstruct the semantic decoupling method by using the graph label representation. Extensive experiments on three benchmark datasets well demonstrate that our proposed framework can signifi-cantly achieve the state-of-the-art performance. In addition, a series of ablative studies further demon-strate the positive impacts of our proposed model.(c) 2022 Elsevier B.V. All rights reserved.