Research On Semantic Image Retrieval And Classification

Semantic image retrieval and classification is a hot research issue in multimedia information retrieval, and is a challenging problem. It has received increasingly widespread concern. Because of the complexity of image semantics, it is difficult to extract semantics, represent it and use it.In the early years, the texture-based image retrieval has had two difficulties. One is that it spends much time and effort on artificial annotation. Another is it is subjective for human to annotate. In order to avoiding these, content-based image retrieval (CBIR) has been developed in 1990s'. It extracts low-level image features automatically, calculates the similarity between images in feature space, and obtains the retrieval results. The CBIR developed fast and became the mainstream technologies in image retrieval.However, because there exists huge difference between low-level image features and human understanding to image, the low-level features cannot describe image content exactly. That is, there are "semantic gap" between low-level features and image semantics. Thus, the semantic image retrieval and classification is proposed. It combines the semantic information of images with visual features to retrieval or classify images. The focus on this technology is how to extract, represent, and use the semantic information.This thesis discusses the key techniques on semantic image retrieval and classification, including semantic extraction, semantic representation, semantic image retrieval and classification.The main contributions of this thesis are as follows:1. Propose a particle swarm optimization-based semantic image classification approach. This approach applies the proposed fuzzy particle swarm optimization clustering algorithm to clustering images into semantic classes. In this approach, it introduced fuzzy into particle swarm optimization in that the image semantics are usually fuzzy or uncertain. It also adopts feature selection method to avoid "dimension disaster" problem, which caused by high dimension of image features.2. Present an image semantic objects recognition method based on rough-fuzzy set. This method first constructs a fuzzy data cube using image low-level features.Then applies presented rough-fuzzy method on this fuzzy data cube to segment image into semantic regions. A new dependent function on rough-fuzzy set is defined. And our approach is different from the traditional rough set-based methods, which are usually used to process discrete data, in that it can process continuous data.3. Develop a semantic image retrieval method using multiple instance learning. Image retrieval is transform into a multiple-instance learning problem. This approach first extracts image simple semantics by multiple instance learning, and then mapping the simple semantics to complex semantics. Because of the characteristics of multiple-instance learning that it only needs to know the label of the image, the image regions in training set don't need to annotation beforehand. This decreases the dependence of the approach to the result of the image segmentation, and simplifies the pretreatment.4. Put forward a semantic hierarchical model to describe image semantics. This model expresses the image semantics as image level, image region level, simple semantics level, and complex semantics level according to the granularity of image semantics. The lower level is mapped to the higher level through certain mapping method. This model describes different granularity of image semantics by difference level. Moreover, it gives a guidance to extract the complex semantics, which is difficult to extract directly.