8.6 Statistics of Natural Scenes

It is often argued [636] that vision is not possible if the physical world does not have enough regularity, because visual tasks are ambiguous and the underlying inverse problems are ill posed [765]. Although certain regularities can be found by experiment [308, 416, 879], the source of any given regularity of natural scenes can be difficult to identify [48, 813]. Most regularities of natural scenes discovered so far are statistical and therefore cannot be applied to any individual scene. However, many practical problems (e.g., the color and density balance problems in image printing and display) do not currently have a reliable, deterministic solution. Most color image processing algorithms today rely on some types of Bayesian estimation and prior statistics are used either explicitly or implicitly. It is thus useful for us to learn about some of the regularities of natural scenes. The important thing to keep in mind here is that, before the cause or the source of a regularity is identified, we should use it with great care. It is always a good practice to build in some error checking steps in the algorithms when using statistical assumptions.