Natural Images

In September 1998, the UK Applied Vision Association held its first meeting on the topic of `̀ Natural Images''. The meeting, which was held at University of Bristol attracted seven submitted papers, together with an invited lecture by David Tolhurst. In discussion, the idea arose that the topics of the papers might be suitable for a Special Issue of Perception. Here is the Special Issue, after a suitable gestation period. There was sustained interest in the topic, with a further meeting in 1999 attracting nine papers, and a third meeting in September 2000, as an adjunct to the British Machine Vision Conference, also attracting nine papers. Why this interest in natural images? There seem to be several reasons. First, it is clear that the human visual system normally deals with images of greater complexity then are traditionally used in psychophysical or physiological studies. The reason for using simplified stimuli is (a) that they are easier to generate, and (b) that experimental control is easier. However, there has been considerable progress, both in computer technology/software, and also in the underlying mathematics. This means that such complex images can now be used as stimuli in a controlled manner and with a good idea of what the information content of the images actually is. The main aim of this issue is to give the reader an idea of how such experimentation can proceed. This also entails looking at the underlying theoretical issues, and these are dealt with in most of the papers in rather different ways. There is, of course, a problem. By the time the reader receives this issue the work is somewhat out of date, and in a rapidly changing field recent developments are therefore not covered. It will be apparent, for example, that the majority of scenes in these papers are monochrome and staticöwhich is clearly not the case for real-world scenes. These topics have been aired in later meetings, but the present issue presents a selection of topics which are still very much relevant. Interested readers are encouraged to submit material and to attend the Natural Images Symposium at ECVP 2001, which is the successor to the AVA meeting.. Soöhow can experiments be carried out in this field? The first question concerns the independent variable: the information content of the images. How can this be succinctly summarised? A common approach to this has been to look at the secondorder statistics of the image. Typically, these are represented by the power spectrum and the phase spectrum. Until recently, the power spectrum has been the most popular measure of information content due primarily to its conceptual tractability. However, as some of the present papers demonstrate, this is changing. While the power spectrum remains an important consideration, the phase spectrum is receiving more attention. This phase spectrum is a complicated variable, and to tease out an understanding of what it is, there is an increasing tendency to tease it apart by considering `higher-order' statistics. The second problem in carrying out research in this field is to know what hypotheses to adopt and how to test them. A way of beginning to tackle this is to suppose that human vision is somehow optimised to encode the critically important information in natural scenes. All that is needed is a method of varying the information content, and a dependent measure which allows an assessment of the efficiency with which an observer can carry out a visual task. The first paper in this issue, by Willmore, Watters, and Tolhurst, provides a good introduction to the issue of what the independent Guest editorial Perception, 2000, volume 29, pages 1013 ^ 1015