Image Scene Classification Methods Research Based On Probability Generated Topic Models

Image scene classification is essential to image analysis and image understanding. With the development of information technology, the amount of image data has achieved rapid growth. It has become one of the important tasks of the current that these massive image data are rapid, accurate classification. Image scene classification is the process of making images automatically for different categories based on some prior knowledge. Image scene classification provides semantic foundation for subsequent processing image, and has been successfully applied to many fields.Image scene classification method is varied, and probability generated topic model is a hot spot of research. The nature of probability generated topic model is representing the image with the vector of topic probability, and the topic is a hidden variable which has certain semantic information. The model expresses the image with bag-of-words firstly, then the model assigns a topic o each word of image, in the result of a topic frequency vector of every image, finally, predicts the image categories using the vector. This paper mainly studies how to determine the topic of each visual word so that the model can achieve better classification effect. The paper improves the classification effect based on the problem of some existing probability generated topic modes such as Latent Dirichlet Allocation, shared component topic models, and has obtained certain research results. The main works are as follows:(1) Shared component topic models have been introduced to analyze the produce process of the topic, which is sampled from one same prior probability. But the probability of each topic of actual image is not equal, which can be deduced easily from analysis images. This paper proposes a hierarchical shared component topic model, by introducing a hierarchical dirichlet process, which can train the prior probability of topic to a vector similar to the actual probability of the topic of image, so that the model can better describe the diversification of the image content.(2) Based on hierarchical shared component topic model, further study of the prior probability of topic has been made, which finds that the prior probability in different categories of image is not equal, so this paper proposes a marked hierarchical shared component topic model. This model is an improved supervised hierarchical shared component topic model, which adds the category tag information of images and estimates the prior probability of topic according to the image category of the corresponding word belong to. The model well describes the differences between image categories.(3) For the difference and consistency between image categories, this paper puts forward the concept of discriminative topic space, which contains a lot of subspace, which is special topic space of each image category and the topic space of all image categories. The paper puts forward a discriminative hierarchical shared component topic model, which introduces a linear transformation to transform the topic of one word to the topic of the discriminative topic space, and end up with a better more compact topic vector to present image.