Graph Convolutional Networks-based Label Distribution Learning for Image Classification

The one-hot vector employed for true label representation has been widely applied for image classification. However, the one-hot representation assumes that a single label is only associated with one instance, which is not reasonable because labels are generally not completely independent and instances may relate to multiple labels for the real scenarios. Such one-hot representation may ignore the relevance among labels that provide more supervision information for model training. To capture and explore such significant relevance in image classification, we propose, in this paper, GCNLDL, a Graph Convolutional Network based Label Distribution Learning approach for classification. GCNLDL builds a directed graph over the images, where each node (sample) is represented by the embedding of an image, where GCN learns the relevance between an input image and multiple training images from different classes to obtain the label distribution vector. The resulting vector is further combined with the one-hot label to recover a realistic label distribution of the input image, which is employed to train the state-of-the-art classification models. Furthermore, a multilayer perception is proposed to learn an effective label correlation matrix to guide information propagation among the nodes in GCN. GCNLDL is capable of capturing the relevance among labels by representation learning graph structure among image samples during training process and produces a better label distribution to guide the training of the state-of-the-art image classification models, resulting in a performance improvement of image classification. Rigorous experimental results on four public image classification datasets show that GCNLDL outperforms other approaches and effectively improves the performance of deep learning classification based models.