Discriminative semi-supervised learning via deep and dictionary representation for image classification

Abstract

Supervised dictionary learning and deep learning have achieved promising performance in the classification task. However, in many real-world applications there usually exist very limited labeled training samples, although abundant unlabeled data is relatively easy to collect. How to effectively exploit the discrimination of unlabeled data is still an open question, hence semi-supervised learning has attracted much attention from wide fields. Semi-supervised deep feature learning has well exploited the feature discrimination from only the discriminative viewpoint, while dictionary representation-based classification has also been applied to semi-supervised learning but with shallow features. In this paper, we propose a novel discriminative semi-supervised learning via deep and dictionary representation (DSSLDDR), which jointly utilizes the discrimination of dictionary representation for data reconstruction and the distinguishing feature of each sample. To exploit the powerful discrimination of dictionary representation, class-specific dictionaries are required to discriminatively reconstruct a sample, with the reconstruction error to predict the sample’s class label. To exploit the semantic information, the deep neural network extracts discriminative features by using multiple nonlinear transformations to generate the powerful descriptor. Then the class-specific dictionary learning and deep network learning are integrated together to conduct more accurate class estimation for unlabeled data and learn a more discriminative classifier, where an entropy regularization is designed to balance and control the class estimation of unlabeled data. Furthermore, we propose the DSSLDDR++, the extension model of DSSLDDR based on consistency/contrastive learning to further improve the accuracy of class estimation for unlabeled data, making a more powerful semi-supervised learning classifier. Extensive experiments on benchmark datasets show the effectiveness of the proposed methods.