Vision is the main theme of the following essay, including an outline of some of the human visual systems' workings, as well as an explanation of the attempts to reproduce these faculties by way of computational models.
Also to be considered are the meaning ; difference of the terms sense ; perception, with respect to visual perception ; computer vision. Following this the statement from the title will be considered.Sensation ; Perception; what's the difference?Briefly, sense is the faculty of receiving sensation, a part of subjective consciousness. Perception is the action of perceiving; which is to become aware of, through the senses.
It could be said that sense is the stage prior to the process of perception, sensation being the early awareness (of a stimulus) & perception being the recognition, processing, & understanding (of a stimulus), usually with respect to that which is held in memory.More specifically, sensation requires a stimulus that is powerful enough to set off a reaction in the nervous system, as a nerve fires at a specific rate according to its threshold. The sense organ/receptor, in this case the eye, detects the stimulus & changes it into an electrical signal, which is conducted to the visual cortex within the brain via the optic nerve. Once the signal arrives in the visual cortex it initiates its own unique firing pattern, the recognition of which brings about the perception of a specific sensory stimulus. There are various theories of perception & generally it is made plainly obvious that sensation & perception are very different:"Perception is not determined simply by stimulus patterns; rather it is a dynamic searching for the best interpretation of the available data" (Gregory, 1966)Gregory was, & is, very influential in theories of perception; he sees perception as an indirect, top down process, whereas Gibson, (1966) sees it as a direct, bottom up procedure, neither view is sufficient on its own to explain the complexity of perception. The Gestalt theory is different again; perception being viewed as an organisational procedure, but this is insufficient to fully explain the process of perception, although it does take into account so called rules of grouping; proximity, similarity, symmetry & continuation, & is concordant with the concept of emergent features in that the whole is greater than the sum of its parts.
Gregory's interpretation then, would take into account not just the current stimulus, but would endeavour to match it to information already stored from past experience, or prior knowledge, but, it seems that we need little or no conscious effort to figure out the (often) rapidly changing scene we perceive. So, how then can this most complex of faculties be implemented in what is known as computer vision ; even with the numerous facilities that have been replicated computationally, can a computer be said to have the capability of true visual perception?Visual Perception.Modern theories of vision began back in 1604 with the description of the formation of the retinal image within the eye by Kepler, followed by the observation of the retinal image by Scheiner in 1619. This was a good start but, with this progress more questions were raised; the image is in two dimensions, how do we see it in depth? At a rough estimate 90% of the information we receive concerning the world around us reaches us through our eyes (Gross, 1998) and our visual processing areas take up a large part of the brain. There are many elements encompassed by the term visual perception.
These include depth ; colour perception, feature, ; boundary or edge detection, object, gesture ; face recognition. The phenomenon of perceptual constancy; that the perception of an object remains static, regardless of changes in size, light ; orientation is another important aspect of visual perception.The pupil controls the amount of light that enters the eye through the iris, the lens then focuses light on the retina (this image is upside down) which has three main layers, rods ; cones, biopolar cells ; ganglion cells. Cones, sensitive to light from three wavelengths (red, green, blue) ; rods, less specialised, responding to light over a wide range of frequencies. The biopolar cells are connected to the rods, cones ; ganglion cells. The ganglion cells form the beginning of the optic nerve leading to the brain.
This is where it gets complicated, with 1/2 the information from each eye going to the opposite side of the brain, through the thalamus to the visual cortices. There are different types of cortical cells, some that respond to specific features, others that respond in a less specialized way.Where the optic nerve connects to the visual cortex there are no rods ; cones, this gives us a blind spot in each eye, but because they don't overlap, information from one eye fills in the information lacking in the other. This blind spot is not noticeable without conscious & deliberate effort, Dennett, (1991) explains this by saying that as there are no parts of the visual cortex responsible to receive information from this area, this lack of information is not noticed.The ability to see in depth derives from the slight difference between the two images presented to the retina, known as stereopsis. These are binocular cues, there are other cues to depth; monocular cues, perspective, relative size, texture & movement.
Familiar size is one cue to distance; our prior experience with objects leads us to suppose that where they appear smaller than others, those which seem to be smaller are expected to be further away, it is similar with relative size. Shadow is another important depth cue. Linear perspective is a kind of depth cue useful in a static, two-dimensional image such as a drawing or photograph.One explanation for colour perception is the Trichromatic theory, which claims that three different types of cone each respond to light from separate parts of the visible spectrum, this still holds today although there is an alternative; the Opponent colour theory, which says that colour vision is reliant on actions by two types of receptor, each reacting in one of two ways. There is evidence to support both theories but a complete theory should include components of each theory.Visual features also inform & enhance other forms of perception such as language, examples of these are kinesic (regarding movement) & standing features (regarding appearance) both being recognised as part of the many non-linguistic cues in human communication.
Pinker (1998) in agreement with Marr, says that vision has evolved to solve the "ill posed problems" that are created by the inverse optics of the human eye, by adding premises, making assumptions about how this world is, on average, put together, some of these are innate, developed over many years of evolution, others are learned through a lifetime.There are also cultural & environmental influences on visual perception; research by Segall, Campbell & Herskovitz (1996) showed that the Muller-Lyer illusion is either missing or lessened in some people. Their research involved Zulus living in circular huts that had arched doors, who had no knowledge of rectangular buildings & weren't affected by the illusion to the same degree as westerners are. They argued because these people live in a 'circular culture' ; are therefore less susceptible to the illusion than those who live in a world of right angles. From this it seems that conventions concerning depth that are taken for granted by westerners are learned, Cole ; Scribner (1974)Computer Models of Some Aspects of Vision.Marr's Computational Theory of Vision.
The initial responses of the retinal cells & of an electronic camera effectively contain a two dimensional array of intensity values, this 1st stage of vision known as the grey level array. The 2nd Stage of vision involves locating changes in intensity; analysis of the grey level array to determine between regions of different intensity. These differences represent boundaries, which could be edges or just changes in reflection & illumination. A mathematical process known as convolving is used to reduce interference from 'noise' along with visual filtering & a comparison of the filtered images yields a high level of information.The 3rd stage of vision is known as the primal sketch, this consists of primitive symbols, such as bars, edges & blobs.
Each has a specific position, orientation, length, width & contrast in intensity with the surrounding region. The symbolic description of this is attached to the map of the filter of the grey level array. All of this, the raw data for the primal sketch roughly shows the complete organisation of the image, & is constructed by the grouping of similar elements again & again. Although this theory has lost some influence as a cognitive model, due to a lack of plausibility in the 3 dimensional sketch, Bruce, et al, (1997) Marr's work was a great influence on the design of computer vision systems.Object Recognition.
Initial work in pattern recognition was often concentrated on alphanumeric characters ; success in this area has been applied to automated postal sorting systems, but it has little relevance to object recognition as they are two dimensional. Theories such as template matching are suited to computational modelling if the stimuli is standardised, but this does not lead to a complete account of how the variations of the three dimensional world is understood. Feature analysis would seem at first to be more plausible than the idea of having a template for everything we can see within our memory, ; the Pandemonium system, Selfridge, (1959) had success with letter recognition, but it is simply an extension of template theory ; as such cannot account for complex object recognition in humans although it has some use computationally, Neisser (1967)Marr ; Nishihara, (1978) developed the idea of prototypes, as recognition of an object is possible from many different angles; they contended that the shape must be specified in co-ordinates established by the object itself. They compiled a catalogue representing complex objects in a hierarchical organisation, the higher levels representing the gross structure, the lower levels making the specific details explicit.
Hogg, (1983) expanded this idea to cope with movement, quite successfully. These approaches to recognition use high level information regarding the shape of objects, employing prior knowledge to interpret the world, this prior knowledge can give back what is lost when lower level cues are unclear. Prior to this, Wallach ; O'Connell, (1953) used an approach called the kinetic depth effect to recover structure from movement.A more recent model with superior credibility than others is that of Biederman, (1987) this theory follows on from Marr's, but objects are seen as a specific grouping of their constituent parts, these parts called geons. He proposes that geons have distinct properties that are invariant regardless of the view, each Geon having its own basic characteristics in the primal sketch representation, from which object recognition can be attained, which rejects the level of the three dimensional sketch.
This gives weight to the validity of this model, as was said before there seems to be a lack of reliability in the three dimensional sketch.Gesture Recognition.Various attempts have been made to implement hand gesture recognition, which is useful for computer games ; in manufacturing for controlling machinery. There has also been research into the recognition of sign language, a difficult area, as not only hand gestures have to be taken into account; but also object ; face recognition, both of which are important to the understanding of sign language. There has been some success with recognition of the manual aspect of sign language, Sagawa, et al (1996) but less success with the grammars ; other previously mentioned features.
Interactive Activation ; Competition (IAC) Model, Burton, Bruce ; Johnston (1990).This connectionist model could be seen as a further development of Bruce ; Young's (1986) identification 'route' model which used incoming structural code of a face which was matched against an array of stored face recognition units, if any of these is activated the face is acknowledged as familiar, this is followed with semantic information which leads to a name being retrieved. There is support for the basic structure of this model but it has limitations. IAC models the later, more cognitive phases of person identification which attempts to address these limitations.
Other, later research has proposed that face recognition may be based on a more complete representation where the whole configuration of the face could be as significant as any element. Bruce, Burton & Hancock (1999) have recently linked the IAC model with a holistic image based addition, which gives a complete model of face recognition.Conclusions.Returning to the initial definitions of the terms sense & perception, as far as having the ability to receive sensation, it is a facility that has been implemented in computers, although, as a computer does not have consciousness, it is difficult to liken this facility to the human concept of sensation. It could be said that, step by step, as computational models evolve to encompass greater complexity & are able to contain more contextual information, they are becoming closer to emulating the wide range of human visual perception, but that doesn't mean that they perceive the world in the manner that humans do.
Again returning to the initial definitions, perception is the recognition, processing, ; understanding of a stimulus, relating to the prior knowledge that is held in memory. We know that a computer can be programmed to recognise, process ; identify a stimulus, the operative word here is 'programmed' as the computer is only carrying out these operations within the constraints of a limited system. Any prior knowledge that is contained within the computer has been placed there ; is not something that the computer has experienced; although 'learning' has been built in with the advent of connectionism ; parallel distributed processing.Returning to attempts to develop a model for gesture recognition, particularly with regard to sign language, the wide array of visual information that needs to be simultaneously processed causes problems because there are so many elements that need to be simultaneously combined in order to present a complete understanding.
It is this lack of readily available contextual information that can confound computer vision systems.A computer does not have a sense of its place within the world ; expectations of what should be where, whereas humans have their lifetime of experience to exploit, with massive inferencing capability using these memories ; prior knowledge. Harnad, (2003) talks about the feeling/function problem; that even if a robot can be built that is impossible to differentiate from humans in what it can do, that the difference is that the robot does not feel, it only behaves, in fact behaves as if it can feel, without actually feeling anything.
