Image Classification by a Probabilistic Model Learned from Imperfect Training Data on the Web

Current approaches to image classification require training images prepared by hand. In this paper, we describe experiments on image classification using images gathered from the Web automatically as training images. To gather images from the Web, we use the probabilistic method we proposed before. In the method, we build a generative model which is based on the Gaussian mixture model (GMM) from imperfect training images gathered from the Web in order to distinguish relevant images from irrelevant images. In this paper, we propose applying the model built during Web image gathering process to generic image classification task. In the experiments, we classified Corel images with the probabilistic model learned from Web images automatically.