Image categorization of integrated generative models and discriminative methods

In our approach, the global visual vocabulary which is similar to keypoints of codebooks is built with Gaussian Mixture Models based on local image features. Images are represented as a new set of feature vectors which are summed posteriori responsibility relative to different visual words. The discriminative classifier is trained by Support Vector Machine with linear kernels based on above features. Experiments were performed on the PASCAL VOC 2006 dataset and the results suggested the influence of background factors on classification effectiveness. And further experiments showed that the features extracted from object areas can be combined effectively to improve classification performance in our method.