Research On Semantic Topic Model Based Image Scene Classification

ABSTRACT:Scene classification annotates automatically images with a group of semantic labels, such as coast, mountain, street, city, forest and so on. It helps to provide effective context for image understanding on high layer. From the perspective of cognitive psychology, the work has been done with the hierarchical structure, following the routine of "Bag of Visterrns—Semantic Topic Model—Scene Classification". The main key point lies in how to train the computer to understand the semantic content of image, and recognize similarity and difference among images.In order to deal with the semantic gap between the low layer and high layer in image description, our work focuses on modeling the image scene, and achieves the following research results:Since images with the same scene category share semantic similarities, our work proposes a topic model based on the category-constrained learning way to capture the specific semantic. According to the different method to construct the topic simplex, the model can be implemented in two ways. One is CTS-LDA (Category Topic Simplex-Latent Dirichlet Allocation), which consists of category models. Because the topic set for each category model is same size, CTS-LDA is easy to extend. The other is ATS-LDA (Adaptive Topic Size-Latent Dirichlet Allocation), whose size of topic is adjusted adaptively based on the content of each category scene. After computing marginal distribution for the image in all category models, the category label of the image can be decided by Maximum Likelihood Selection in our model.Our work proposes another topic model to learn image scene category. In order to solve the local extremum problem in EM, we take advantage of the role of pseudo-count in the Dirichlet parameter, and execute Variational Inference two times to deduce model parameters. In the first inference, the prior distribution of general topics is estimated, which is used as the initial value for the prior distribution of category topics in the second inference. The inference way of our model is easy to implement incremental learning. For the new coming scene category, the model can learn its prior distribution parameters based on the prior distribution of general topics learned before. As a result, our model exhibits better generalization ability.Our work proposes a topic model with spatial semantic, which can capture the con-occurrence information among visterms. At first, the model builds a pyramid to preserve the spatial context for features by fusing local features and global features, based on original Bag of Visterms. Then, it takes into full consideration the common of general topic and uniqueness of category topic on the middle layer. At last, it builds the semantic topic simplex contained spatial context based on general topic and category topic. Therefore, the image semantic description represented in the topic simplex of our model shows more discriminability than prior methods.