Abstract
In this paper, we propose an image classification method using local features and linear manifolds. In our method, local features such as SIFT are extracted from test and training images. In each class, all local features of training images are compressed by principal component analysis, thus a set of local features extracted from training images is represented by a linear manifold (affine subspace). In a voting phase, dissimilarity between a local feature extracted from a test image and each linear manifold is measured, and one vote is added to the class to which the nearest linear manifold belongs. Through voting by all test local features, the test image is then classified into the class that has majority votes. In this paper, we also adopt a steepest descent based learning algorithm to our classification rule for improving accuracy. The experimental results on the WANG image dataset verifies that our method performs as well or better than a nearest neighbor voting method using all training local features.
