Research Of Middle-level Semantic Based Image Scene Classification Algorithm

With the rapid development of multimedia and information and technology, it is also rapid growth in image data. Because of vast amounts of data, the effective management in computer has become an urgent task. For solving the problem, we propose image scene classification, which has large role in image retrieval, object recognition, computer vision and other fields.Image scene classification system based on image content information can divide images into categories automatically, such as coast, forest, city, and other scenes. Currently, the method of image scene classification consists of based on underlying-feature and based on middle-feature. Between the underlying image feature and high-level semantic has gap, so based on the underlying feature can't achieve effectively image scene classification. This paper describes from the beginning of image underlying-feature, around of established the middle semantics, at last completing image scene classification.Middle semantics is based on visual dictionary. The step of creating the visual dictionary is:the first extracting low-level visual features, and then forming visual words through K-means clustering. Finally these visual words constitute visual dictionary. The description of the image transforms the underlying feature into visual words. As the problem of synonymy and polysemy in dictionary, we reference the ideas of LSA in text statistics. We firstly establish the theme model, and with probabilistic latent semantic analysis (PLSA) identify potential themes. We use the maximum likelihood model to complete the scene classification of image. However, this method directly classify from the appearance of image overall situation. It does not consider the spatial distribution of the image. In view of this, on the basis of original method we divide images into some block. Each block as a whole image, we extract potential theme in PLSA. Finally, we combine the potential theme vector of each block. These vectors are as the SVM input vector to complete scene classification of image. Experiment show that introduction of supervised SVM increases the complexity, but this method completed scene classification of image more accurately.