Research On Image Semantic Classification Techniques In Content-Based Image Retrieval

Image semantic classification is an important and challenging task in the field of Content-Based Image Retrieval (CBIR). Since digital images and videos are becoming a major source of multimedia data, grouping images into semantically meaningful categories is a very imperious demand. Traditional techniques of CBIR try to retrieve images through analyzing the similarity of image visual features, but CBIR cannot meet the requirements of semantic image retrieval. Classifying image database into reasonable categories using low-level image features, will greatly improve the performance of CBIR systems.Firstly this thesis introduces the contents, architectures, and several practical systems of classical CBIR. The extraction techniques of visual image features are specially discussed as research fundamental, and these techniques include extraction of color, shape, texture, and spatial-relationship features etc. The primary techniques of image semantic classification are also mentioned. Discussions about image semantic retrieval are given with the analyses of CBIR's drawbacks.Image semantic model is an abstract of image's semantic representation and process, and it also provides the reasonable research directions. Bayesian probability framework is a theory to transfer a priori probability to posterior probability. This thesis provides a formal Bayesian framework for image classification problems, which maps the low-level image features to the intrinsic high-level semantics.There are a variety of representations for global image features, and they are all the important bases of image classification research. An image semantic classification approach based on single global image feature is proposed. A most-discriminating single global feature is selected based on the theory of relevant feedback. At the same time, this algorithm adopts the image division strategy from conventional photography theory. Classifiers are implemented for Indoor/Outdoor and City/Landscape classification problems, and intensive analysis of the algorithm is given with the experimental results on large image databases.Multiple global image features can also be incorporated together to categorize images. The advantages of using multiple global image features is discussed, and a novel image semantic classification algorithm is proposed based on the integration of multiple global image features andspace distribution information of features. A new feature representation is proposed. And an incremental learning scheme is adopted to improve the classification performance. Experimental results show the new algorithm is especially suitable for indoor/outdoor classification problem that has relatively stable space distribution of features.The local image features are always corresponding to specific semantic classes. Active Appearance Models algorithm (AAM) is utilized to describe similar appearances of objects in images that fall into the same class because of the similar objects. Aiming at the familiar human faces and red-eyes detection problems, experiments are performed on real household digital album. Results show that the approach has very high retrieval accuracy.Finally, a prototype system is developed based on this thesis's research. Some details of implementation are discussed and the practicability of this thesis's research is demonstrated.