Attention-based label consistency for semi-supervised deep learning based image classification

Abstract

Semi-supervised deep learning, which aims to effectively use the available unlabeled data to aid the model in learning from labeled data, is a hot topic recently. To effectively employ the abundant unlabeled data and handle the imbalance in labeled data, we propose a novel attention-based label consistency (ALC) model for semi-supervised deep learning. The relationships between different samples are well exploited by the proposed scheme of channel and sample attention; meanwhile, the class estimations are required to be smooth for nearby unlabeled data. The proposed ALC is further extended to the imbalanced case by developing a label-imbalance ALC model. We have implemented the proposed ALC model in the semi-supervised frameworks of Π model and MeanTeacher, and the experimental results on four benchmark datasets, (e.g., Fashion-MNIST, CIFAR-10, SVHN, and ImageNet) clearly show the advantages of our proposed method.