The Research And Application Of Image Classification Based On Topic Model And Graph Kernel

Image classification plays an important role in the field of computer vision especially in biological data analysis and military traffic study. With the boom of machine learning, image classification techniques have developed rapidly and become a hot research field.Topic model is an emerging field in machine learning algorithms and it opens up a new frontier for image classification research. Topic model is a probabilistic generative model which is mainly used in text information processing. Latent Dirichlet Allocation (LDA) model is spring up based on the model PLSA which is widely used in text classification. The main idea of LDA is based on bag of words model and every document is regard as a vector of word frequency. Furthermore, it transforms the text information to the digital information. With the development of topic model, LDA has been successfully applied in image classification and image retrieval.However, the idea of bag of words regards that the documents(images) consist of a list of disordered words ignoring the spatial structure of the image. Graph kernel is a tool of computing the similarity of graph structure. In our work, we combine topic model and graph kernel to process image from the sight of semantic and spatial respectively. The main work in this paper is as follows:1) This paper addresses semantic image annotation, focusing on developing hidden topics in image with traditional LDA model. Firstly, Scale-invariant feature transform(SIFT) algorithm is available to detect feature points and K-means is used to form the dictionary of visual word. We use the Gibbs sampling for parameter estimation and then predict the new images. Secondly, LDA model successfully reflect the high level features and the RGB SIFT features which integrating the SIFT features with color features.2) In this paper, graph kernel is studied and applied in image classification successfully. We take into account several common graph kernels and propose to use the minimum spanning tree algorithm kernel to calculate the similarity between images based on the merits of it. Firstly, Support vector machine(SVM) is available for classification with the similarity matrix. We propose to utilize the minimum spanning tree kernel as the kernel function of support vector machine. Secondly, the weight of nodes is taken into account while calculating the similarity matrix of graph structure. The experimental evaluations the COREL database demonstrate its promise of the effectiveness.3) We combine the improved LDA model and graph kernel according to the strengths and weaknesses of them. In order to get the similarity matrix of pictures, cosine similarity is used after the improved LDA model. Furthermore, this similarity matrix is mixed together with the similarity matrix of graph kernel under a certain scale. Finally, we use SVM to implement the images classification. The new model fully considers the color information, semantic information and spatial information of images in image classification. Experiments show that our method achieves comparable results compared with some traditional approaches.