If accurately searching for documents across the internet – or an intranet – is difficult, then how much more difficult must it be to quickly find an image – a picture of a horse in a field, perhaps, or a well-known personality? If the picture is labeled, then the job is certainly much easier. But if it is not? And what if the user is not even sure exactly what he or she is looking for – a witness helping the police look for a suspect, for example. Unsurprisingly, many leading universities have been attracted to the challenge to devise an acceptable method of image recognition – especially one that can be used across the Internet. There are projects underway at dozens of leading universities, including MIT, Stanford, Iowa, Columbia, and Berkeley in the US, and at Cambridge in England and Edinburgh University in Scotland. Several large computer companies, notably IBM, have carried out advanced research in this area. And, of course, there is the secret work carried out on behalf of the Department of Defense. It is interested in how a computer can quickly spot a pattern or a shape – say a fast moving aircraft – and know exactly what it is seeing. Researchers at Iowa have even developed a program which can, up to a point, find pictures of naked people from the web. This clearly has huge commercial possibilities.
Massive task
How do these work? The technology is similar to that of searching for a keyword in a document. But instead of looking for a string of characters, the searcher submits a pattern – for example, a black cross against a white background. A first stage will often be to trawl through databases looking for files with image ‘tags’, such as ‘gif’ or ‘tif’. Then the processor’s task of matching the patterns is not technically difficult – although it can be massive, since image files are often vast in size. The program will usually have a store of special images which it can recognize, such as the shape of ‘limbs’ or facial characteristics. In a sense, it needs to understand context, in the sense that a fingerprint should not be confused with a relief map. A key issue is how to frame accurate searches using a special interface. For example, the program may need to accept simple drawings, or it may allow users to pick colors from a palate. Several young companies have sprung up in the image recognition area, many of them set up by university graduates seeking to commercialize project work. These include Excalibur, perhaps the best known company developing image searching software. Its Retrievalware software is being used by, among others, Yahoo! and Computer Associates, and several major financial and government organizations. Interpix and Virage, the internet multimedia and image searching companies, are also growing fast. Virage supplies a component for facial recognition which can be plugged into the Oracle database. Text specialist Autonomy of Cambridge also has some image capability, although it has not so far commercialized its work. According to Meta Group analyst Clive Longbottom, the lack of image searching capability is a big issue for text retrieval companies such as Fulcrum and Verity.
