This chapter shows how to use text mining techniques to search for images based on their visual content. The basic ideas of using visual words are presented and the details of a complete setup are explained and tested on an example image data set.
Content-based image retrieval (CBIR) deals with the problem of retrieving visually similar images from a (large) database of images. This can be images with similar color, similar textures, or similar objects or scenes: basically any information contained in the images themselves.
For high-level queries, like finding similar objects, it is not feasible to do a full comparison (for example using feature matching) between a query image and all images in the database. It would simply take too much time to return any results if the database is large. In the last couple of years, researchers have successfully introduced techniques from the world of text mining for CBIR problems, making it possible to search millions of images for similar content.
The vector space model is a model for representing and searching text documents. As we will see, it can be applied to essentially any kind of objects, including images. The name comes from the fact that text documents are represented with vectors that are histograms of the word frequencies in the text. In other words, the vector will contain the number of occurrences of every word (at the ...