Research Of Image Content Representation And Multi-label Annotation Algorithm

With the increasingly advent of multimedia technology and Internet, the number of im-ages has been explosively increased. An emerging issue is how to effectively manage and retrieve such a huge resource of digital images. Since the early1900’s, Content-Based Image Retrieval(CBIR) has replaced Text-Based Image Retrieval(TBIR) and becomes a new research hotspot in this subject. However, the well-known "semantic gap" restricts the developmen-t of CBIR. So researchers take their attention back to semantic, and Semantic-Based Image Retrieval(SBIR) becomes a central issue.In this paper, with in-depth understanding and analysis of image semantic retrieval depend-ing on the latest research, we innovatively put forward the RoI-BoW model for representation of image content, as well as the image annotation model based on content representation with multi-layer segmentation(MLSIA).RoI-BoW is improved on the basis of BoW, which considers the importance of RoI for image retrieval. Firstly, RoI is achieved by key point detection and filtration. Secondly, RoI and Non-RoI are represented by different methods respectively. Finally, the two parts are com-bined together to represent image content. Representation of image content based on RoI-BoW is applied to do experiments on image retrieval, and compares with representation of image content based on BoW. Experimental results prove that retrieval with representation of image content based on RoI-BoW can get more accurate results.Employing multi-layer segmentation for representation of image content and annotating images with second-order CRFs are the main innovations of MLSIA. Firstly, multi-layer seg-mentation is used to represent image content, which combines saliency analysis and normalized cut(Ncut), and utilizes region-based BoW to represent image content. Then MLSIA inputs rep-resentation of image content and semantic labels to train second-order CRFs, and uses the trained CRFs to annotate images. Experimental results show that MLSIA can achieve promis-ing performance for multi-labeling, and outperform the model based on single-layer segmenta-tion.