Technical: SD Sampling for Recognition
The great significance of Nyquist sampling is that provides an exact copy of the input optical image. Nyquist sampling should therefore be used for vision systems whose function is to generate an output image for display to a human operator. However for vision systems whose function is to perform a task, such as recognition, there is no requirement for an output image, and therefore no specific need for Nyquist sampling. In fact Nyquist sampling is actually unsuitable for most recognition tasks because it generates a mass of irrelevant and obstructive point-by-point detail. SD sampling should be used for vision systems whose function is to perform a visual task such as recognition.
The great significance of SD sampling is that almost all recognition by human vision is performed using SD information derived by the peripheral retina (99.99% of the total retina). An object (e.g. tank, tree, truck, car, etc) is instantly recognized in a single glance, in which almost the entire area of the object is necessarily SD-sampled by the peripheral retina because the very small fovea (0.01% of the retina) is over only one very small part of the object during the short period of the glance. (To positively identify the object a very small key part of it may subsequently be specifically fixated and Nyquist- sampled by the fovea). Reading is an extremely important human visual task in which, once again, almost all of the visual information used is derived by the peripheral retina - that is by SD sampling. If each individual character of text had to be foveally fixated reading would be extremely slow and intolerably tedious.
The primary justification for the proposed use of SD sampling for recognition is that human vision, which has immensely superior recognition capability, uses visual information derived almost entirely by SD sampling. However it is particularly noteworthy that bio-inspired SD sampling is, quite independently, also justified by a first-principles analysis of the recognition problem.
The fundamental problem of automatic recognition is that the image of the object to be recognized will almost always differ in point-by-point detail from the reference pattern that defines it (because of pose differences, design differences, damage, etc, etc). The object may therefore not be recognized by a system based upon high density Nyquist sampling. However low-density SD sampling provides high reolution information need for recognition yet enables the commonality of these different views to be recognized because of its low pixel density.
Adequately high resolution is essential for recognition by any means. Nyquist sampling provides this high resolution by its high sample-density. However a mass of point-by-point detail is thereby inevitably generated (by the high sampling density) which will not correspond between different views of the same object, and in particular between the sensed image of an object and the reference pattern that defines it. The mass of point-by-point detail generated by Nyquist sampling is not only irrelevant but actually highly obstructive of recognition. Nyquist sampling is inherently unsuitable for automatic recognition!
On the other hand low pixel density SD sampling captures high-resolution information essential for recognition without a mass of non-corresponding point-by-point detail. SD sampling is inherently suitable for automatic recognition! With low pixel density SD sampling the target image and its reference pattern will correspond much more closely and the object can then be reliably recognized in spite of pose and other differences.