Semi-supervised image classification based on sparse coding spatial pyramid matching

Image classification, namely classifying thousands of images into different classes, is an important task in images organization. Although many existing methods attempt to address this task, most of those are proposed in a supervised way based on the labeled data. However, in real world the labeled data is usually hard to obtain while large amounts of unlabeled data can be easier to acquire. The problem of effectively and efficiently classifying images combining unlabeled data with labeled data remains pretty much open. To this end, in this paper we proposed a novel semi-supervised image classification method based on sparse coding spatial pyramid matching (ScSPM). Specifically, we use the unsupervised ScSPM method to get the representation of unlabeled images as like the labeled images. Based on the obtained image representation, we then propose a linear LapSVM as the semi-supervised classifier. Since the proposed method has a linear kernel and can effectively explore the intrinsic structure of data by making full use of the information of unlabeled data, it leads to more accurate and efficient image classification. Experimental results on two real world datasets demonstrate the effectiveness of our method especially when the labeled data is very little.