Image Classification Based On The Bag-of-Words Model

The performance of the bag-of-words model in the field of image classification is mainly limited by the quantization error of the local feature. In this thesis an image classification method based on global coding combined with multi-scale codebook is proposed in order to reduce the quantization error of the local feature effectively. We implement global coding by making full use of the manifold structure of image features and computing the global information of the codebook. The coding coefficients obtained by our method are fairly smooth and accurate. Furthermore we design a multi-path method to integrate all feature representations to describe the image. To a certain extent this method can achieve the scale invariance of feature representations. We conduct several experiments on three commonly used benchmark data sets UIUC-8、Scene-15 and Catltech-101, and the average classification accuracy rates reach up to 88.0%、83.9%�'� 83.1% respectively. The experimental results show that our method significantly improves the performance compared with the fixed-scale locality-constrained coding methods.The main contributions of this thesis are as follows:Firstly a multi-scale feature learning method based on multi-scale codebook is proposed. The multi-scale codebook based on the spatial structure of the bag-of-words model is used to improve the ability to describe of codebook and reduce the quantization error of the local feature. Based on the multi-scale codebook we design a multipath coding method to process the extracted feature to obtain multi-scale feature representation in order to obtain all levels of information of the image.Secondly we propose a global coding method to make full use of multi-scale codebook. The global information is added in the coding process to convert local coding into global coding. In this way we further reduce the quantization error to make feature coding more smooth and make feature representation more accurate. Due to the robustness of global coding, we do not need to find the optimal values of K to obtain the optimal classification effect.Finally we combine the multi-scale feature learning method with the global coding method and propose a multi-scale global coding method for image classification. We analyzed the complexity and factors of our method through the experiment on several benchmark data sets. The experimental results prove the effectiveness of our method. At last we discuss the pros and cons of the method.