Research On Image Classification Based On Region Of Interest

With the quantity of image data growing rapidly, image classification is becominga hot spot of research interest. It enables users to browse and retrieve images in astraightforward manner, and is the premises for efficient image indexing and processing.Although low-level feature abstraction techniques in image processing has beenmatured, higher levels of feature abstraction such as to recognize what objects the imagecontains and the spatial relations between these objects are still at its primitive stage.The current image processing and computer vision technologies are unable to solve thisproblem so far. As a result, automatic semantic classification of image database is still adifficult research subject and hasn't been applied in today's image database products.This thesis studies the previous research results and made an effort to solve theproblems of image content analysis, feature abstraction, classification based on SVM.The work done in this thesis and its main progresses includes the followings:1. The design of image classification framework and the research on techniquesused in image classification. A thorough research on visual saliency from theperspective of computer vision was done, and based on that the semantic informationcontained in an image was analyzed, and a distinction on the main classificationinformation in an image--objects that can represent the semantics of theinformation—was made. An image classification framework is proposed. The nature ofimage classification is pattern recognition, and according to its theories, an imageclassification procedure is designed. Also key techniques involved are also discussed.2. The research on feature selection strategies of interested region. A region ofinterest detection algorithm is studied based on an image classification framework withregard to region of interest. In the meantime, to ensure the stability and robustness ofimage feature abstraction, SIFT descriptors are thoroughly studied in the field of imagematching, and are applied in image classification field. A feature selection strategy isproposed after SIFT extraction, and the effectiveness of features is discussed.3. The research on SVM construction on the basis Matching Kernel. A thoroughstudy of SVM theory is carried out. Based on the characteristics of features fromregions of interest, a Kernel Function is adopted and improved to lay down a sound basefor the design and implementation of SVM classifier. And utilizing this Kemel Function,a classifier aimed at image classification problem based on areas of interest isconstructed. In practical, the use of two-class classifier is limited, and based on"one-against-one" theory, a multi-class classifier is designed, and validated on ObjectCategories image database.