Image Retrieval From The Web & Archives

The World Wide Web allows us to access masses of information, both textual and visual. Conducting a search for images by entering a few keywords into a search engine is simple, but such a search often results in hundreds and sometimes thousands of images being returned. Many of these images will be totally unrelated to the subject of the search as current image searching is largely based on words rather than image content.

Oxford Inventors collaborating with colleagues at California Institute of Technology have now produced a system to solve these challenges. The new technology involves searching by image, sometimes called content based image retrieval (CBIR), rather than word, and gives a much more effective result.

The Oxford Invention

Representation, detection and learning are the crucial steps in designing a visual system for recognising object categories. The first challenge is to capture the common features of an object category, but at the same time being sufficiently flexible to allow for variability in shape, lighting, viewpoint etc. The challenge of detection is defining the metrics and inventing the algorithms that can match models to images in the presence of background clutter. Effortless learning is the ultimate challenge, and requires that the training sets should be small with operator assisted steps ideally eliminated.