Discriminative Semi-supervised Learning Based on Visual Concept-Like Features

Abstract

A discriminative semi-supervised learning method based on visual concept-like high-level features is proposed in this paper. Previous semi-supervised learning methods usually use unlabeled data to augment the training set or regularize the decision boundary of classifiers. The classification results rely on the precision on unlabeled data using supervised classifiers trained with limited labeled samples. When a small number of labeled samples are provided, these methods are likely to get bad results. Differently, the proposed method directly uses the distribution information of all available data in the feature space to learn a new representation which is achieved by computing the similarities of a chosen image and some discriminative data exemplars sampled from the feature space. A semi-supervised distance metric learning method by learning a projection matrix under the equivalence constraints of similar pairs and dissimilar pairs is introduced to measure these similarities, and a pseudo-mahalanobis distance is thus obtained to represent the similarities between data samples instead of Euclidean distance. Experiments showed the effectiveness of this learned distance. The new representation can be fed into standard classifiers for image classification task. The training data of our system can either be original image data or handcrafted features or image features learned by deep architectures. Therefore, the proposed method can be applied in both feature extraction and feature enhancement. In the semi-supervised classification task on eight standard datasets, the proposed method achieves improved performance over many of the previous existing methods.