Affordable Access

Transfer Component Analysis for Domain Adaptation in Image Classification

Publication Date


Transfer Component Analysis for Domain Adaptation in Image Classification Giona Matascia, Michele Volpia, Devis Tuiab,c, Mikhail Kanevskia a IGAR, University of Lausanne, Baˆtiment Amphipoˆle, 1015 Lausanne (Switzerland); {giona.matasci, michele.volpi, [email protected] b IPL, University of Vale`ncia, 46100 Burjassot-Vale`ncia (Spain); c LASIG, Ecole Polytechnique Fe´de´rale de Lausanne, 1015 Lausanne (Switzerland); [email protected] ABSTRACT This contribution studies a feature extraction technique aiming at reducing differences between domains in image classification. The purpose is to find a common feature space between labeled samples issued from a source image and test samples belonging to a related target image. The presented approach, Transfer Component Analysis, finds a transformation matrix performing a joint mapping of the two domains by minimizing a probability distribution distance measure, the Maximum Mean Discrepancy criterion. When predicting on a target image, such a projection allows to apply a supervised classifier trained exclusively on labeled source pixels mapped in this common latent subspace. Promising results are observed on a urban scene captured by a hyperspectral image. The experiments reveal improvements with respect to a standard classification model built on the original source image and other feature extraction techniques. Keywords: Domain adaptation, Feature extraction, Transfer Component Analysis, Image classification. 1. INTRODUCTION In remote sensing image classification the ground truth collection process can be very demanding. Therefore, when classifying series of similar images with the supervised learning paradigm, the possibility to reuse labeled samples from a first acquisition is very appealing. Particularly, the ability to adapt a classifier built on an image, the source domain, to a new scene without needing any (or needing little) labeled data from the second image, the target domain, is of remarkable interest.1 In

There are no comments yet on this publication. Be the first to share your thoughts.