Manifold-Ranking-Based Keyword Propagation for Image Retrieval
- PDF / 454,338 Bytes
- 10 Pages / 600.03 x 792 pts Page_size
- 69 Downloads / 212 Views
Manifold-Ranking-Based Keyword Propagation for Image Retrieval Hanghang Tong,1 Jingrui He,1 Mingjing Li,2 Wei-Ying Ma,2 Hong-Jiang Zhang,2 and Changshui Zhang1 1 Department 2 Microsoft
of Automation, Tsinghua University, Beijing 100084, China Research Asia, 49 Zhichun Road, Beijing 100080, China
Received 30 August 2004; Revised 29 January 2005; Accepted 5 April 2005 A novel keyword propagation method is proposed for image retrieval based on a recently developed manifold-ranking algorithm. In contrast to existing methods which train a binary classifier for each keyword, our keyword model is constructed in a straightforward manner by exploring the relationship among all images in the feature space in the learning stage. In relevance feedback, the feedback information can be naturally incorporated to refine the retrieval result by additional propagation processes. In order to speed up the convergence of the query concept, we adopt two active learning schemes to select images during relevance feedback. Furthermore, by means of keyword model update, the system can be self-improved constantly. The updating procedure can be performed online during relevance feedback without extra offline training. Systematic experiments on a general-purpose image database consisting of 5 000 Corel images validate the effectiveness of the proposed method. Copyright © 2006 Hindawi Publishing Corporation. All rights reserved.
1.
INTRODUCTION
The initial image retrieval methods are based on keyword annotation and can be traced back to the 1970’s [1, 2]. In such approaches, images are first annotated manually with keywords, and then retrieved by their annotations. As long as the annotation is accurate and complete, keywords can accurately represent the semantics of images. However, it suffers from several main difficulties, for example, the large amount of manual labor required to annotate the whole database, and the inconsistency among different annotators in perceiving the same image [3]. Moreover, although it is possible to extract keywords for Web images from their surrounding text, such extraction might be far from accurate and complete [4]. To overcome these difficulties, an alternative scheme, content-based image retrieval (CBIR) was proposed in the early 1990’s, which makes use of low-level image features instead of the keyword features to represent images, such as color [5–7], texture [8–10], and shape [11, 12]. Its advantage over keyword-based image retrieval lies in the fact that feature extraction can be performed automatically and the image’s own content is always consistent [4]. Despite the great deal of research work dedicated to the exploration of an ideal descriptor for image content, its performance is far from satisfactory due to the well-known gap between visual features and semantic concepts, that is, images of dissimilar semantic content may share some common low-level features, while
images of similar semantic content may be scattered in the feature space [4]. In order to narrow or bridge the gap, a great deal of work has bee
Data Loading...
