Publication

Abstract : In the past few years in computer vision and machine learning, transfer learning has become an emerging research for leveraging richly labeled data in the source domain to construct a robust and accurate classifier for the target domain. Recent work on transfer learning has focused on learning shared feature representations by minimizing marginal and conditional distributions between domains for linear data sets only. However, they produce poor results if they deal with non-linear data sets. Therefore, in this paper, we put forward a novel framework called Kernelized Transfer Feature Learning on Manifold (KTFLM). KTFLM aims to align statistical differences and preserve the intrinsic geometric structure between the labeled source domain data and unlabeled target domain data. More specifically, we consider Maximum Mean Discrepancy for statistical alignment and Laplacian Regularization term for incorporating manifold structure. We experimented using benchmark data sets such as the PIE Face Recognition and the Office-Caltech (DeCAF features) object recognition dataset to discourse the limitations of the existing classical machine learning and domain adaptation methods. The performance comparison indicates that our model gave splendid accuracy of 79.41% and 91.97% for PIE and Office-Caltech data sets using linear and Gaussian kernels, respectively.