Image Classification by Labeled Data in Similar Topics

One of the most difficulties in image classification is the short availability of labeled images. However, in many cases, we can find a large amount of auxiliary labeled images to help the target classification. The auxiliary images are in similar topics of the target images. For example,donkey may be used as auxiliary data of horse, since they are quite similar with each other. In this paper, we propose to use similar but different labeled images as auxiliary data to help train image classifiers. An SVM-based classification algorithm is proposed to tackle our problem. Our algorithm first maximizes the margin between the decision boundary and the training instances, and then minimizes the penalty corresponding to misclassification. We conduct extensive experiments to demonstrate the effectiveness of our algorithm. The experimental results show that our algorithm can greatly improve the image classification performance by making use of similar labeled image as auxiliary data, against several state-of-the-art image classification algorithms.