An investigation of trained neural networks from a neurophysiological perspective.

The application of theoretical neural networks to preprocessed images was investigated with the aim of developing a computational recognition system. The neural networks were trained by means of a back-propagation algorithm, to respond selectively to computer-generated bars and edges. The receptive fields of the trained networks were then mapped, in terms of both their synaptic weights and their responses to spot stimuli. There was a direct relationship between the pattern of weights on the inputs to the hidden units (the units in the intermediate layer between the input and the output units), and their receptive field as mapped by spot stimuli. This relationship was not sustained at the level of the output units in that their spot-mapped responses failed to correspond either with the weights of the connections from the hidden units to the output units, or with a qualitative analysis of the networks. Part of this discrepancy may be ascribed to the output function used in the back-propagation algorithm.

Color and spatial structure in natural scenes.

Digitized records of terrain scenes were produced using a technique of photographic colorimetry. Each record consisted of three tristimulus images (X, Y, and Z) which were analyzed for their color statistics, spatial frequency content, and image correlation. Interactions between color and space were examined using a cone receptor transformation. It is shown that the scene amplitude spectra follow an approximate reciprocal variation with frequency, and that the correlation function can be described by a one-step autoregressive model. The results are discussed in terms of methods for optimum image coding in human and machine vision.

A self-similar stack model for human and machine vision.

A new model is proposed that not only exhibits the major properties of primate spatial vision but also has a structure that can be implemented efficiently in a machine vision system. The model is based on a self-similar stack structure with a spatial resolution that varies with eccentricity. It correctly reproduces the visual cortical mapping function, yet it has the important attribute that it can produce invariant responses to local changes in the size and position of image features. By proposing a novel purpose for cortical "bar-detectors", the model can also produce invariance to more general distortions. The structure of the model allows efficient hierarchical search to be made and it naturally embraces the concept of "attention area". Exploitation of this model has already confirmed these properties and has also revealed its robust ability to control the focus and gain of machine vision systems.

Discrimination and detection thresholds: the effect of observer criterion on the spatial properties of chromatic and achromatic mechanisms.

Spectral sensitivity functions were determined for structured test targets presented on a white background field. Data obtained for threshold detection of the targets are consistent with results obtained from previous studies which used unstructured test fields showing functions, with three maxima, dependent on opponent colour mechanisms. Data obtained from threshold discrimination measurements show a marked reduction in the blue sensitivity when 1 degree test targets are used. This effect decreases significantly when larger targets are used. A change of the observer's criterion can significantly alter the relative contributions of the achromatic and chromatic channels at threshold.

Visual form perception and the spatial phase transfer function.

The quality of many imaging devices can be characterized, within certain constraints, by means of the modulation transfer function (MTF) and the phase transfer function (PTF). In many cases, it is possible to estimate, qualitatively, the effect of the MTF on the appearance of objects, and much progress has been made in making quantitative predictions of the detectability of objects and features within objects. This is not the case, however, for the PTF, and its influence is often neglected, even though nonideal PTF's obviously may degrade image quality. Experiments are described that attempt to assess the significance of the PTF for human visual performance. The effects of various PTF's were simulated by means of a technique that maintained the modulation of the spatial-frequency components closely constant. The visual detectability of such phase changes was measured by a discrimination method, and phase threshold levels were evaluated. Simple linear patterns provided the targets for all measurements, although some images are shown that illustrate qualitatively the phase-shift results on real two-dimensional targets. The work enables an assessment to be made of the effects of phase changes produced by various imaging systems.