Search

Support for the hypothesis that the early visual experience of adventitiously blind persons gives them a visual frame of reference after the onset of blindness is found in studies of tactual form discrimination and finger maze learning and in experiments involving “extended space.” The length of early visual experience is an important factor and is related to the progressive establishment of cross-modal correspondences in the early months of life.

We outline a theoretical framework that leads from the computational nature of early vision to algorithms for solving them and finally to a specific class of analogue and parallel hardware for the efficient solution of these algorithms. The common computational structure of many early vision problems is that they are mathematically ill-posed in the sense of Hadamard. Regularization analysis can be used to solve them in terms of variational principles of a specific type that enforce constraints derived from a physical analysis of the problem. Studies of human perception may reveal whether principles of a similar type are exploited by biological vision. We also show that the corresponding variational principles can be implemented in a natural way by analogue networks. Specific electrical and chemical networks for localizing edges and computing visual motion are derived. We suggest that local circuits of neurons may exploit this unconventional model of computation.

Descriptions of physical properties of visible surfaces, such as their distance and the presence of edges, must be recovered from the primary image data. Computational vision aims to understand how such descriptions can be obtained from inherently ambiguous and noisy data. A recent development in this field sees early vision as a set of ill-posed problems, which can be solved by the use of regularization methods. These lead to algorithms and parallel analog circuits that can solve 'ill-posed problems' and which are suggestive of neural equivalents in the brain.

Fifty years worth of data and dozens of doctoral students later, Franklin M. Henry remembers his early vision of physical education and how the field evolved into a modern academic discipline.

I review a new theoretical framework that from the computational nature of early vision leads to algorithms for solving them and suggests a specific class of appropriate hardware. The common computational structure of many early vision problems is that they are mathematically ill-posed in the sense of Hadamard. Standard regularization analysis can be used to solve them in terms of variational principles that enforce constraints derived from a physical analysis of the problem, see T. Poggio and V. Torre (Artificial Intelligence Lab. Memo No. 773, MIT, Cambridge, Mass., 1984). Studies of human perception may reveal whether some principles of a similar type are exploited by biological vision. It can also be shown that the corresponding variational principles are implemented in a natural way by analog networks, see T. Poggio and C. Koch (Artificial Intelligence Lab. Memo No. 783, MIT, Cambridge, Mass., 1984). Specific electrical and chemical networks for localizing edges and computing visual motion are derived. These results suggest that local circuits of neurons may exploit this unconventional model of computation.

EVE, theEarly VisionEmulation software, is a set of computer programs designed to compute models of early visual processing. EVE may be used with a wide variety of models concerning spatial detection and discrimination, motion analysis, and issues of spatial sampling. EVE is modular and flexible. It runs under the UNIX operating system, and is device-independent. We describe the implementation of the EVE software and discuss how it may be applied to several visual models.

For six years after the invention of the transistor in 1948, the development lay commercially neglected. The U.S. military was interested in applying it, but most commercial and consumer engineers resisted any changeover from vacuum tubes. At Texas Instruments in Dallas, the story was different. Patrick Eugene Haggerty, then the company's executive vice president, had caught an early vision of the transistor's potential, and it helped alter the course of commercial electronics.

Several computational theories of early visual processing, such as Marr's zero-crossing theory, are biologically motivated and based largely on the well-known difference of Gaussians (DOG) receptive-field model of retinal processing. We examine the physiological relevance of the DOG, particularly in the light of evidence indicating significant spatiotemporal inseparability in the behaviour of retinal cell types. From the form of the inseparability we find that commonly accepted functional interpretations of retinal processing based on the DOG, such as the Laplacian of a Gaussian and zero crossings, are not valid for time-varying images. In contrast to current machine-vision approaches, which attempt to separate form and motion information at an early stage, it appears that this is not the case in biological systems. It is further shown that the qualitative form of this inseparability provides a convenient precursor to the extraction of both form and motion information. We show the construction of efficient mechanisms for the extraction of orientation and two-dimensional normal velocity through the use of a hierarchical computational framework. The resultant mechanisms are well localized in space-time and can be easily tuned to various degrees of orientation and speed specificity. Plusieurs theories computationnelles du traitement de la vision primaire, comme la theorie du croisementzero de Marr, sont biologiquement motivees et basees largement sur le celebre modele du champ receptif de la difference de gaussiens (DDG) applique1 au traitement retinien. Nous examinons ici la pertinence physiologique de la DDG, particulierement en tenant compte de ľevidence indiquant une inseparabilite spatio-temporelle significative dans le comportement de la cellule retinienne. A partir des caracteristiques de ľinseparabilite, nous trouvons que les interpretations fonctionnelles classiques du traitement retinien basees sur la DDG, tels que le laplacien ?un gaussien et les croisementszero, ne sont pas valides pour les images variant dans le temps. Contrairement aux approches actuelles en vision automatique, qui tendent a separer dans une premiere etape ľinformation sur la forme de celle sur le mouvement, il apparait que les systemes biologiques fonctionnent tres differemment. On montrera plus loin que la qualitye de cette inseparabilite donne un premier outil approprie pour ľextraction de ľinformation portant a la fois sur la forme et sur le mouvement. Nous decrivons la construction de mecanismes efficients pour ľextraction de ľorientation et de la vitesse 2-d normale a-travers ľutilisation ?un cadre computationnel hierarchique. Les mecanismes resultants sont bien localises dans ľespace-temps et peuvent aisement ětre regies selon divers degres de specification de ľorientation et de la vitesse. Mots cles: vision biologique et automatique, difference de gaussiens, images variant dans le temps, extraction de la vitesse, traitement hierarchique.

Mathematical results on ill-posed and ill-conditioned problems are reviewed and the formal aspects of regularization theory in the linear case are introduced. Specific topics in early vision and their regularization are then analyzed rigorously, characterizing existence, uniqueness, and stability of solutions. A fundamental difficulty that arises in almost every vision problem is scale, that is, the resolution at which to operate. Methods that have been proposed to deal with the problem include scale-space techniques that consider the behavior of the result across a continuum of scales. From the point of view of regulation theory, the concept of scale is related quite directly to the regularization parameter lambda . It suggested that methods used to obtained the optimal value of lambda may provide, either directly or after suitable modification, the optimal scale associated with the specific instance of certain problems. >

Natural images are noisy and, as a result, locally ambiguous. As the tasks associated with low-level vision are spatially local, these ambiguities represent one of the major problems in providing useful intrinsic representations. Numerous techniques exist for symbolically coding images so as to minimise local ambiguity by only recording ‘significant’ changes in image contrast. These techniques invariably employ models, often in the form of global filters, that produce local edge maps. Unfortunately, these methods typically suffer in that they produce hesitant edge maps that often contain severe feature distortions. We propose an alternative approach whose goal is the recovery of underlying surface. An adaptive Surface Labelling technique (ASL) is employed which dynamically places best-fit surfaces to image patches on the basis of their local noise statistics. The methods has been implemented and demonstrates an ability to remove considerable local ambiguity whilst offering near perfect feature preservation.

Computer algorithms have been developed for several early vision processes, such as edge detection, stereopsis, motion, texture, and color, that give separate cues to the distance from the viewer of three-dimensional surfaces, their shape, and their material properties. Not surprisingly, biological vision systems still greatly outperform computer vision programs. One of the keys to the reliability, flexibility, and robustness of biological vision systems is their ability to integrate several visual cues. A computational technique for integrating different visual cues has now been developed and implemented with encouraging results on a parallel supercomputer.

Background: Poor follow-up after cataract surgery in developing countries makes assessment of operative quality uncertain. We aimed to assess two strategies to measure visual outcome: recording the visual acuity of all patients 3 or fewer days postoperatively (early postoperative assessment), and recording that of only those patients who returned for the final follow-up examination after 40 or more days without additional prompting. Methods: Each of 40 centres in ten countries in Asia, Africa, and Latin America recruited 40–120 consecutive surgical cataract patients. Operative-eye best-corrected visual acuity and uncorrected visual acuity were recorded before surgery, 3 or fewer days postoperatively, and 40 or more days postoperatively. Clinics logged whether each patient had returned for the final follow-up examination without additional prompting, had to be actively encouraged to return, or had to be examined at home. Visual outcome for each centre was defined as the proportion of patients with uncorrected visual acuity of 6/18 or better minus the proportion with uncorrected visual acuity of 6/60 or worse, and was calculated for each participating hospital with results from the early assessment of all patients and the late assessment of only those returning unprompted, with results from the final follow-up assessment for all patients used as the standard. Findings: Of 3708 participants, 3441 (93%) had final follow-up vision data recorded 40 or more days after surgery, 1831 of whom (51% of the 3581 total participants for whom mode of follow-up was recorded) had returned to the clinic without additional prompting. Visual outcome by hospital from early postoperative and final follow-up assessment for all patients were highly correlated (Spearman's rs=0·74, p

Human psychophysics and monkey physiology studies have shown that attention modulates early vision – contrast sensitivity and processing. But how can we bridge the effects of attention on perceptual performance to their neural underpinnings? Here we implement a population-coding model that estimates attentional effects on population contrast response given psychophysical data. Model results show that whereas endogenous (sustained, voluntary) attention changes population contrast-response via contrast gain, exogenous (transient, involuntary) attention changes population contrast-response via response gain.

This is the first study to demonstrate the selectivity of learning for contrast polarity. The finding is the main result of an investigation into the existence of central and peripheral vision mechanisms selective for contrast polarity within the texture-segregation process, using the perceptual learning paradigm in a detection task. Energy models (Malik & Perona, 1990) exclude segregation of textures composed of elements of odd-symmetric luminance profile by contrast polarity differences. Here the target was a Gabor patch (0.8 deg) of 1cyc/deg in sine phase (odd-symmetry) embedded in a background of mirror-image elements. Our results showed that, in fovea, segregation on the basis of contrast polarity was above threshold from the first session. After learning, the target popped-out in both central and peripheral vision for durations over 10ms. Our major result is that learning is selective for contrast polarity; it is also selective for orientation and position, all characteristics distinctive of early processing. Since the learning effects were obtained with texture composed of odd-symmetric mirror-image elements, they indicate that the output from odd-symmetric filters was not excluded or inhibited in texture segmentation, but instead played an active role. Our data support models of texture segmentation, in which detection of texture gradient is achieved on the basis of early cortical process, before the non-linear transformation of their output.

Visual rehabilitation practices at a centre for visually handicapped children in Sweden are described and a new vision test, specifically designed for the young visually and/or multiply impaired child, is presented.

SummaryA recent study provides new insights into how the very different response characteristics of the main visual pathways from the eye to the brain may directly result from the presence or absence of a ‘spike trigger-zone’ in retinal bipolar cells.

Convolution, one of the most widely used functions in early vision processing, is, at the same time, highly demanding in terms of computing power. This paper describes a dedicated coprocessor able to perform ID and 2D convolutions; the latter case is applicable only if the convolving bidimension- al filter is separable. Details concerning the main building blocks are given together with experimental results showing the overall performances of the system.

We present an isotropic neural network model for processing by layer IVc of the primate primary visual cortex. It describes how this layer can reconstruct fine local details in an image which have been lost in the low-capacity retinal-LGN pathway, while at the same time narrowing the effective spatial-frequency bandwidth of the response to sinusoidal patterns. We also investigate the circumstances under which such a network can act as an elementary feature extractor by responding preferentially to striped, checked or other high-symmetry patterns. We find that the model can act in something like this way in a particular region of its parameter space, but that such behaviour is incompatible with the reconstruction of local detail.

A. I. Laboratory Working Papers are produced for internal circulation and contain proteins, lipids, cholesterol, polysorbate-80, and other compounds unsuitable for external exposure. It is not intended that material in this paper be applied externally; it is intended for internal consumption only. Serving suggestion: add taco sauce (not included)., I outline a project for integrating several early visual modalities based on coupled Markov Random Fields models of the physical processes underlying image formation, such as depth, albedo and orientation of surfaces. The key ideas are: a) to use as input data estimates of the various processes and their discontinuities, computed by several different algorithms. b) to implement with MRFs the physical and geometrical constraints of local "continuity" of the processes and of their discontinuities. Processes are coupled to each other: the most common form of coupling is a veto — one process vetoing another — as in the case of discontinuities and the associated continuous field., MIT Artificial Intelligence Laboratory

A quarter of a century ago I introduced two paradigms into psychology which in the intervening years have had a direct impact on the psychobiology of early vision and an indirect one on artificial intelligence (AI or machine vision). The first, the computer-generated random-dot stereogram (RDS) paradigm (Julesz, 1960) at its very inception posed a strategic question both for AI and neurophysiology. The finding that stereoscopic depth perception (stereopsis) is possible without the many enigmatic cues of monocular form recognition - as assumed previously - demonstrated that stereopsis with its basic problem of finding matches between corresponding random aggregates of dots in the left and right visual fields became ripe for modeling. Indeed, the binocular matching problem of stereopsis opened up an entire field of study, eventually leading to the computational models of David Marr (1982) and his coworkers. The fusion of RDS had an even greater impact on neurophysiologists - including Hubel and Wiesel (1962) - who realized that stereopsis must occur at an early stage, and can be studied easier than form perception. This insight recently culminated in the studies by Gian Poggio (1984) who found binocular-disparity - tuned neurons in the input stage to the visual cortex (layer IVB in V1) in the monkey that were selectively triggered by dynamic RDS. Thus the first paradigm led to a strategic insight: that with stereoscopic vision there is no camouflage, and as such was advantageous for our primate ancestors to evolve the cortical machinery of stereoscopic vision to capture camouflaged prey (insects) at a standstill. Amazingly, although stereopsis evolved relatively late in primates, it captured the very input stages of the visual cortex. (For a detailed review, see Julesz, 1986a)

I review a new theoretical framework that from the computational nature of early vision leads to algorithms for solving them and suggests a specific class of appropriate hardware. The common computational structure of many early vision problems is that they are mathematically ill-posed in the sense of Hadamard. Standard regularization analysis can be used to solve them in terms of variational principles that enforce constraints derived from a physical analysis of the problem, see T. Poggio and V. Torre (Artificial Intelligence Lab. Memo No. 773, MIT, Cambridge, Mass., 1984). Studies of human perception may reveal whether some principles of a similar type are exploited by biological vision. It can also be shown that the corresponding variational principles are implemented in a natural way by analog networks, see T. Poggio and C. Koch (Artificial Intelligence Lab. Memo No. 783, MIT, Cambridge, Mass., 1984). Specific electrical and chemical networks for localizing edges and computing visual motion are derived. These results suggest that local circuits of neurons may exploit this unconventional model of computation. © 1985 Academic Press, Inc.

This paper describes attempts to model the modules of early vision in terms of minimizing energy functions, in particular energy functions allowing discontinuities in the solution. It examines the success of using Hopfield-style analog networks for solving such problems. Finally it discusses the limitations of the energy function approach.

The problem of symbolic representation of intensity variations in gray-value pictures of real scenes is studied. The goal is to relate the responses of a filter bank of different gradient filters to the structure of the picture which is determined by the physics of the image generation process. A simple criterion is proposed for the selection of a suitable center frequency of the involved band-pass filters. The gradient vectors of the image function give the direction of maximal intensity changes with high resolution (8 bit) which can be used for an invariant shape description by corner points of a contour. The picture is segmented by closed contour lines into regions which form a topographic representation in the picture domain.

The goal of this project was to develop fast algorithms for performing the computations required by the various early vision modules that we have been developing. The early vision modulus require intensive computation and are a major problem in system development since they consume a distractingly large amount of real time during the development process. The goal of the work therefore was high computational efficiency, to be accomplished both by improving the basic efficiency of the algorithms as well as multiprocessing, so that the development of larger systems that use the early vision modulus as components would not be adversely affected by slow response times of the modules.

Under the appropriate conditions, zero-crossings of bandpass signal are very rich in information (Logan, 1977). We examine here the relevance of this result to the early stages of visual information processing, where zero-crossings in the output of independent spatial-frequency-tuned channels may contain sufficient information for most of the subsequent processing.

We outline a theoretical framework that leads from the computational nature of early vision to algorithms for solving them and finally to a specific class of analogue and parallel hardware for the efficient solution of these algorithms. The common computational structure of many early vision problems is that they are mathematically ill-posed in the sense of Hadamard. Regularization analysis can be used to solve them in terms of variational principles of a specific type that enforce constraints derived from a physical analysis of the problem. Studies of human perception may reveal whether principles of a similar type are exploited by biological vision. We also show that the corresponding variational principles can be implemented in a natural way by analogue networks. Specific electrical and chemical networks for localizing edges and computing visual motion are derived. We suggest that local circuits of neurons may exploit this unconventional model of computation.

ME. JOHN HENRY REYNOLDS, whose death occurred on July 17 at tne advanced age of eighty-five years, may be truly described as one of the great pioneers of technical education in Great Britain. Though his work was wrought chiefly in Manchester, his influence was felt throughout the whole of the United Kingdom, and even beyond the seas, and it is not too much to say that he is to be numbered among those to whose early vision and service we are indebted to-day for the great development in the teaching of technology and applied science during the past forty or fifty years. Mr. Reynolds' work began in days when the need for technical education had not been realised, save by an enlightened few, and he steadily set himself to the task of awakening interest in what he knew to be a thing of vital import to the industries of Great Britain—the provision of the highest instruction and training in science and technology for the equipment of those who are to guide and direct and, by the use of special knowledge, develop industrial work. He was in the highest sense an idealist, and in his early outlook visualised a national system of education which would afford a means of consecutive training from the elementary school to the highest work of the university for students of proved ability and application, however humble their circumstances. That he lived to see the fulfilment in large measure of his ambitions was due, in part at least, to his own strenuous endeavours and clear vision.