Image Classification Research Based On Bag Of Words Models

Image classification is one of the hot topics in computer vision, and has wide applications in image retrieval, behavior analysis, medical imaging, intelligent search engine, etc. Recently, benefiting from the new algorithms and theories in pattern recognition, machine learning and image feature extraction, significant progress has been made in image classification. However, due to the great variations in images of the same class and the visual appearance similarity between images of different classes, image classification still is facing big challenges. On the basis of Bag of Words models, this thesis delves into image classification using feature fusion, sparse coding and point-to-set distance metric learning.Considering the fact that different features are complementary, this paper jointly uses HSV features and SIFT features to perform image classification.A novel image classification algorithm is proposed combining metric learning together with feature fusion. Specifically, two image representations are adopted that are HSV color histogram and the resulting vector by pooling sparse codes of SIFT features respectively. Based on these two representations, two distance metrics are learned by optimizing point-to-set distances. Compared with other classifiers, the nearest subspace classifier can have better classification performance when coupled with point-to-set metric learning.An image classification method is proposed by use of Fisher vector representation. Fisher vector representation possesses the advantages of both generative models and discriminative models. In addition, compared to traditional BoW models, It can model high order statistics of SIFT features. Instead of encoding SIFT features by sparse coding algorithm and perform pooling, the proposed algorithm makes use of both HSV color histogram and Fisher vector of SIFT features to perform image classification.We have tested our proposed methods on a flower image database. The difficulties lie in great variations in flowers of the same class and similarity between flowers of different classes. Experiments have demonstrated the good performance of proposed methods compared with other state of the art algorithms.