Semi-Supervised Image Classification in Likelihood Space

This paper studies the problem of using limited amount of labeled data and large amount of unlabeled data in the training of a generative model for image classification, and proposes a likelihood space approach to improve the classification performance. Frequently when labeled data is limited, unlabeled data can help to improve classification performance if the assumption of the generative model structure in the classifier is correct. But classification accuracy can be degraded if the model structure assumption is incorrect. In this paper, we compare raw data space classification and likelihood space classification in semi-supervised learning framework, and we show that the classification performance can be improved in likelihood space when model is misspecified. We apply this likelihood space semi-supervised learning method in automatic target recognition on SAR images, and experimental results demonstrate the effectiveness of this proposed approach.