2.2.2. Model for Peripheral Level of Visual System

    The human visual system can be viewed as a cascade of two systems, as shown in Figure 2.14. The first system. which represents the peripheral level of the visual system, converts light to a neural signal. The second system. which represents the central level of the visual system, processes the neural signal to extract information.

    Unlike central level processing, about which little is known, peripheral level processing is fairly well understood, and many attempts have been made to model it.    One very simple model for a monochrome image that is consistent with some well-known visual phenomena is shown in Figure 2.15. In this model. the monochrome image intensity I(x,y) is modified by a nonlinearity, such as a logarithmic operation, that compresses the high level intensities but expands the low level intensities. The result is then filtered by a linear shift-invariant (LSI) system with spatial frequency response H(Ω_x,Ω_y). The nonlinearity is motivated by the results of some psychophysical experiments that will be discussed in the next section. The LSI system H(Ω_x,Ω_y), which is bandpass in character, is motivated by the finite size of the pupil. the resolution limit imposed by a finite number of light-receptive cells, and the lateral inhibition process. . The finite size of the pupil and the resolution limit due to a finite number of light receptive cells are responsible for the lowpass part of the bandpass nature of H(Ω_x,Ω_y). The lateral inhibition process stems from the observation that one neural fiber responds to many cones and rods. The response of the neural fiber is some combination of the signals from the cones and rods. While some cones and rods contribute positively, others contribute negatively (inhibition). This lateral inhibition process is the rationale for the highpass part of the bandpass character of H(Ω_x,Ω_y). Even though the model in Figure 2.15 is very simple and applies only to the peripheral level of the human visual system, it is consistent with some of the visual phenomena .

One way to exploit a model such as the one in Figure 2.15 is to process an image in a domain closer to where vision takes place. This can be useful in some applications. In image coding, for example, the information that is in the image but is discarded by the visual system does not need to be coded. By processing an image in a domain closer to where vision takes place. more emphasis can be placed on what is important to the visual system. This is one reason why some image processing operations are performed in the log intensity domain rather than the intensity domain.

    In addition to the model in Figure 2.15, more sophisticated models for monochrome images and models for color images have also been proposed in the literature.