Transfer learning using adaptive SVM for image classification

Abstract

Transfer learning is a learning paradigm which enables us to transfer knowledge gained in one domain to other familiar domains. These approaches are useful in scenarios where one domain has large amount of labelled data and another domain has either none or very few labelled examples. In this work, we have used feature extraction techniques (such as PCA, SURF and Gabor filter) to implement transfer learning between human face images in the source domain and images of cat faces in the target domain. Specifically, this work focuses on using the adaptive SVM for classification in the target domain. The novelty of this work is characterized by the use of multiple features for transfer learning, which are robust and sensitive to image orientation, texture and shape. Our results indicate effective transfer learning between the source and target domains, based on the fact that the classifier performs better in the target domain as it learns on more examples in the source domain.