I have been involved with Tom Troscianko, Ian Moorhead and Marilyn Gilmore in developing and evaluating a model of the human visual cortex which is intended to allow us to predict whether a human observer would be able to distinguish between pairs of real or simulated photographs of complex natural scenes. Such models could be used to determine the visibility of camouflaged targets, for instance; or they could be used to assess the fidelity of reproduction of photographs by computer graphics hardware or printers.

The model is based on the known neurophysiological properties of simple cells in the visual cortex of experimental animals, and it is an aim of the model to provide a veridical record of real neurophysiological observations. It assumes that each point in a photograph is evaluated by a set of simple cells, whose optimal stimulus orientations and spatial frequencies cover a range of values. There are thus many thousands of model cells, differing in receptive-field location (compared to the picture borders) as well as in orientation and spatial frequency. The responses of the model cells are not just linear dot-products with picture, but include division by the local mean luminance in the picture to give the nonlinearity of contrast responsiveness. In order to determine whether two complex scenes are discriminable, we must calculate the response of each of the thousands of cells to the two pictures under comparison. It is then necessary to determine how much the cells’ responses differ between the pictures, and whether the differences would be large enough to alert an observer to the fact that the pictures were different. We are also interested in whether we can provide a useful metric of how discriminable pictures are when the differences are clearly suprathreshold.

The idea of this model is, by no means, unique and our implementation is probably similar to that of others. But, the several groups will have had to re-invent the same wheels and to face the same assumptions, and it may be that we (as well as other groups) may not always have recognised the assumptions. I hope that, at the campfire, we could search out these assumptions and the solutions in order to determine which ones are critical and which ones are niceties.

The development of our model and some of its kludges and “features” have been closely affected by the particular pairs of pictures that we were asked to compare: the same natural scene photographed at different times of day (with and without shadowing), or the presence or absence of a vehicle in the natural scene. However, we would want the model to be truly versatile and, at the campfire, I would be particularly interested to know whether other participants could suggest other kinds of task for the model to solve and to be tested against ­ other paired comparisons. Perhaps, our model is not versatile enough and its implementation is too closely focussed on a few tasks.

The most difficult and (probably) most contentious part of the model comes when we compare the responses of each simple cell to the two pictures. How do we model whether the paired responses of one simple cell are different enough in magnitude to be detectable? How do we combine the tiny cues provided by each of the many simple cells to provide an overall measure of discriminability?

Usually, in such models a single cell’s responses are thought to be discriminable if they differ by an amount predicted from a curve like the human contrast discrimination curve ­ the “dipper” function. The dipper function is usually thought of as a reflection of a non-linear relation between response magnitude and contrast. It is also usually thought that a single dipper function will describe a cell’s coding, irrespective of whether a particular response comes from a low-contrast optimal stimulus or a high-contrast non-optimal stimulus. We have been studying how the “dipper” function might arise from the population response of cortical neurons whose limited dynamic ranges cover different parts of the contrast scale. When we also model contrast normalisation, which seems to be an important nonlinearity in cortical processing, it seems that the shape of the dipper may vary under different circumstances. At the campfire, I should like to discuss the implications for discriminations models of having to use more complicated discrimination rules than can be modelled by a simple “dipper”.

I would also like to discuss how we should pool small cues from different cells.