Three approaches to study vision:
Here human eye-brain is treated as a piece of equipment whose properties are studied by observing its responses (otuput) to various stimuli (input) under different conditions. Typically the threshold for detecting a certain feature of some visual stimulus is determined and then how this threshold changes as a function of certain attributes of the stimulus is studied.
To get down to the physiological substrate of vision, the neurons involved in vision are individually studied by recording their responses to various visual stimuli presented to their receptive fields (RFs). Neurons selectively responsive to different attributes of the visual stimuli (such as color, orientation, spatial and temporal frequency, motion, etc.) are discovered. In the last few years, modern technologies such as positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) have been widely used to study the human brain in general, and vision in particular, by detecting the activity of local cortical areas (mm resolution) while responding to various (visual) stimuli.
The previous two methods try to discover how the visual system, at both the neuronal and perceptual levels, respond to different visual stimuli. But computational method goes farther in trying to answer the question how. It is not enough to find neurons that are selective to certain visual attributes. Visual perception is the product of the brain which is composed of basically neurons. So we know there must be neurons and population of neurons that have certain property such as selectivities for various attributes of the visual stimuli. However, to really understand the brain, the important question is HOW it does what it does. Computational methods treat vision problem as an operation applied to the input (stimulus) and generating an output (response). In other words, the output is a function of the input and can be therefore determined, or computed (compute: ``to determine especially by mathematical means''- Encyclopedia Britannica) based on the input.
There is a spectrum of different vision computation methods ranging from robot or computer vision at one end to biological vision modeling at the other. While the former is free to use any mathematical and engineering approach to implement an algorithm for a certain vision task (detection, recognition, etc.), the latter is subject to biological realities. An algorithm mathematically elegant may not be biologically plausible.
Some point of views in signal processing have been adopted in vision study to treat vision as a process of visual signal processing. Consequently, various concepts and tools in signal processing have been used in computational vision, such as linear systems, convolution and filtering, Fourier analysis, etc.
Neural computation is another set of tools for vision computation. The main feature of neural computation distinguishing it from the conventional (serial) computation is that computation is carried out in an artificial neural network (ANN) biologically inspired by the neurons in the brain. ANN algorithms were initially developed to solve pattern recognition problems (e.g., character recognition, perceptron, Rosenblatt 1957). They are widely applied to many practical problems as well as modeling the brain activities.