| Annotating images with keywords has received increasing attention in computer vision community in recent years. By establishing the mapping between images and keywords, the organization, analysis and retrieval of large collections of images will become easier. Region-based image annotation, also known as region-naming, region-labeling, and multi-class image segmentation, is one of the most important methods for image annotation, which assigns metadata to image regions automatically and plays an important role in image understanding, interpretation, and retrieval.However, due to the complexity of the image recognition, it’s difficult to readily transform low-level image features to high-level semantics. Recent studies have shown that contextual information, such as co-occurrences and spatial relationships is able to improve the accuracy of image annotation greatly. Graphical models are common ap-proaches employed to model contextual relationships in region-level image annotation.In this paper, we propose several methods to annotating image regions automati-cally with contextual information, the details are as follows:(1) Propose a supervised topic model for region-level image annotation. Co-occurrence is a commonly used type of context to improve annotation performance, and many researchers employ topic models to capture co-occurrences. However, most pre-vious works utilize topic models to apply image-level annotation. This paper presents a supervised topic model csLDA (class-specified Latent Dirichlet Allocation), which can be utilized to do region-level image annotation and is able to capture co-occurrences between keywords and class labels. Experimental results show that, csLDA is able to improve annotation performance comparing with other state-of-art methods.(2) Propose an integrated framework for image segmentation and pixel-level an-notation. Most of the previous image annotation techniques are based on segmented image regions, and the initial segmentation performance greatly influence the anno-tation accuracy. This paper present a pixel-level image segmentation and annotation model SPLSA (Supervised Probability Latent Semantic Analysis). The proposed mod-el is designed to do image segmentation by modeling pixel-level co-occurrences and the segmentation results are used to build annotation models. Meanwhile, the classi-fication results are feedback to the segmentation framework and improve overall seg-mentation performance. Experimental results show that SPLSA is able to improve the segmentation and annotation accuracy mutually.(3) Propose a new graphical model LDA-CRF to combine two different contex-tual relationships:co-occurrence and spatial interactions. The co-occurrences of vi-sual words and class labels are modeled by utilizing LDA. Meanwhile, CRF model is employed to capture the spatial interactions of visual features and topic information between adjacent regions. The experimental results show that combining two different contextual relationships performs better than utilizing one certain contextual informa-tion alone.(4) Propose an efficient graphical model for long range interactions. Most pre-vious spatial models, such as CRF and MRF, only use interactions between adjacent regions for computation efficiency. However, long range interactions play an important role in image annotation. This paper propose an efficient graphical model ASRG (Ap-proximated Supporting Region Graph) to model long range relationships between im-age regions. We introduce supporting regions to describe spatial interactions between different regions and build a directed graphical model by selecting a group of support-ing regions for each image region. Experimental results show that our proposed model achieves good performances in both annotation accuracy and classification speed.
|
PDF Full Text Request