The Study Of Image Classification And Annotation Based On Probabilistic Topic Model

In computer vision, and machine learning, image classification and annotation have been important, and are important methods obtaining image sementic. They have been applied widely, however, they also meet big challenge. Probabilistic graphical model and variational inference is a new machine learning framework, and is a good method applied to uncertain problem or complex problem. It is applied in computer vision and natural language processing widely. Topic model is original from text processing. An advantage of topic model is data dimension reduction, and another one is it can use the latent topic feature for classification, so that it can narrow the distance between category concept and features. The paper considers that the existing image classification methods and annotation methods are limited in the complex real images, and so we propose a series of probalistic topic models for image classification and annotation. We aim to improve the performance of image classification and annotation. In particular, the innovation is shown as follows:1. The paper proposes an emsemble supervised probabilistic topic model. In the real word, the similarity of intro-class is big, the similarity of inter-class is small. In general, only one classification "criterion"(classifier) is difficult to fit the relation between the complex image and class label. For the complex image data, we think it should exist multiple criteria. It is the problem which ensemble methods can resolve. Ensemble methods can combine multiple weak classifiers to construct a strong classifier. Much work based topic model for image classification has been done now. They mostly construct a single classification criterion for all training data. Our motivating intuition is, here, introducing the ensemble classification idea to topic model, so as to construct a model which can combine the merits of the two kinds of methods. The experimental results on two real image datasets show the effectiveness of the proposed model. 2. The paper proposes a multi-view supervised probabilistic topic model. In computer vision, multiple view features are obtained easily. For complex images, a single feature can not have enough recognition. Obviously, multiple features have better recognition capability than a single feature. The better recognition, the easier classification. The existing topic models for classication mostly focus on single view of features. If they try to consider the information of multiple features, they can only use the union feature of multiple features. Generally, however, simple union of multi-features is not reasonable because of different scales among the views. Thus, motivated by this, we try to construct a probabilistic topic model for classifying data with multi-view features. The experimental results show the effectiveness of the proposed model.3. The paper proposes a probabilistic topic model which can impove image annotation using category information. We consider that category information can provide certain evidence, or valuable information for image annotation. Once the category of the images was ascertained, it is equivalent to reduce some uncertainty of annotation. More over, category information is obtained easily comparing with annotation information in computer vision. It motivates us to construct a probabilistic topic model which can impove image annotation using category information. The experimental results show the effectiveness of the proposed model.4. The paper proposes a novel probabilistic topic model for simultaneous image classification and annotation. We draw on the fact that once the category of an image is ascertained, the scope of annotation words for the image can be narrowed. Meanwhile, the probability of generating irrelevant annotation words can be reduced, and vice versa. As such, we think not only these two tasks of image classification and image annotation can be performed simutaneously, but also they can be implemented in ways that improve one another. Based on this intuition, we propose a novel probabilistic topic model for simultaneous image classification and annotation. The performance of the proposed model is demonstrated on two real-word datasets, and the results show that our model provides a competitive classification performance with several benchmark classification models, while it shows a better annotation performance than other benchmark annotation models. The experimental results also show the rationality of the proposed model.