Content-based image retrieval (CBIR) has wide applications in public life. Either from a static image database or from the Web, one can search for a specific image, generally browse to make an interactive choice, and search for a picture to go with a broad story or to illustrate a document. Although CBIR has been well studied, it is still a challenging problem to search for images from a large image database because of the well-acknowledged semantic gap between low-level features and high-level semantic concepts. An alternative solution is to use keyword-based approaches, which usually associate images with keywords by either manually labeling or automatically extracting surrounding text from Web pages. Although such a solution is widely adopted by most existing commercial image search engines, it is not perfect. First, manual annotation, though precise, is expensive and difficult to extend to large-scale databases. Second, automatically extracted surrounding text might by incomplete and ambiguous in describing images, and even more, surrounding text may not be available in some applications. To overcome these problems, automated image annotation is considered as a promising approach in understanding and describing the content of images.
Key Terms in this Chapter
Photo Classification and Browsing: This is designed for private albums. All the photos are classified into several categories corresponding to some topics. When users want to search for a photo or just randomly browse,they can first select an interesting topic.
Nonparametric Density Estimation: If there is no assumption about the distribution, the density is directly estimated from the training data.
Parametric Density Estimation: Assume that the images and text words are in a certain distribution. Their joint distribution is in virtue of a hidden variable. The parameters of the distribution are estimated by the EM algorithm.
Content-Based Image Retrieval (CBIR): CBIR is the process by which one searches for similar images according to the content of the query image, such as color, texture, shape, and so forth.
Automatic Image Annotation: This is the process in which images are associated with symbols (keywords or tags) automatically instead of manually being labeled as in some approaches. Most of the approaches are implemented by machine learning, pattern recognition, and computer vision. The annotated symbols describe the image content in the semantic concept level.
Web Image Search: This is one of the aspects of Web search. The current image search is based on surrounding text. In the future, it is desired to apply content-based search to Web images by automatic image annotation.
Probability Association Approach: According to Bayes’ theorem, the annotated text words for an image are selected by the posterior probability (the conditional probability of the word, given the image).