Research On Image Scene Classification Based On Probabilistic Topic Model

It is a highly-informed and lively age. The information science and technology develop fast. The data saved on internet, multimedia and mobile devices and exchanged among them increase rapidly. The new technology of intelligent device and internet, as an instance, is applied in every aspect of our life. The economy of the society and the quality of life has reached a higher level due to the development of information technology.In all kinds of data, the images have taken a large part and the proportion is rising every day. They are generated by intelligent devices, especially the mobile devices. How to solve the actual problem in life based on image data has become hot. The topic of this article, the scene classification, is worth researching in the image processing field. Meanwhile, the research on the scene classification is difficult because of the complexity and huge amount of the image data.Many methods have been proposed to classify scene images until now. Scene classification based on topic models, such as Latent Semantic Analysis (LSA), Probabilistic Latent Semantic Analysis (PLSA) and Latent Dirichlet Allocation (LDA), are researched a lot in recent years. In this paper, an improved method based on PLSA and LDA model is the key point. At the beginning of the paper, the background, research status and related application of scene classification is introduced. Next is the introduction of the basic principle of PLSA and LDA probabilistic topic models and KNN and SVM classifiers. After that improved models and algorithms are proposed, the classification experiences are done and the experience results are compared in the text. The main research of this paper is as follows:On the one hand, we use the hybrid framework of generative/discriminative approach which combines the unsupervised topic model and supervised classifiers to classify images. Color, spatial and shape features are combined to form the visual words and visual vocabulary in the process of feature extraction. When calculating and quantizing spatial features, an improved method of chain code is used. While color features are first extracted in RGB, HSV and Lab color space respectively in every block and then global color features are merged. The PLSA model, as an intermediate representation, is used to reduce the dimension of the feature vectors, train and test the visual theme distribution of the image. A hybrid classifier called KNN-SVM is proposed to classify the visual theme distribution.On the other hand, an improved LDA model is proposed and applied to scene classification. SIFT features are added in the feature extraction process. They are extracted in every block, merged to form global visual words and fused with color, spatial and shape features. The hybrid classifier KNN-SVM we proposed is used in the classification process.When conducting the experiments of scene classification, the image dataset is divided into two parts. One is used to extract visual words, train the best parameters of PLSA and LDA model and compute the probability distribution of the latent topics. The other is used to test the classification performance of our model. Finally the experiment results are analyzed. The mean classification precision of different features, different classifiers and different topic models are compared. They all influence the classification result. The comparison proves that our model lift the classification performance. The comparison of experiment results can benefit further study on scene classification at the same time.