There is a large gap between academic vision research and industrial machine vision applications. A cross-fertilization between academic and industrial researchers is of mutual benefit: it leads to practical applications of benefits to industry as well as offering real-world challenges that deepen research understanding. A cross-disciplinary atmosphere between academia and industry should be encourages by appropriate funding sources in government, academia, and industry should be encouraged by appropriate funding sources in government, academia, and industry. Three real-life projects in the author's experience at General Motors demonstrate the potential payoffs of such an approach. An understanding of human hyperacuity mechanisms led to a subpixel method for finding circuit board locations to extremely high accuracy using a low-cost machine vision system. Research on the neural basis of primate color vision led to a proposed new type of color display, based on opponent mechanisms, that should be cheaper than conventional RGB displays. Also, a study of primate motion vision led to machines on the plant floor, and the sensor and display devices needed for vehicles of the 21st century, will greatly benefit from closer interaction between academic and industrial researchers.

This report surveys cognitive aspects of color in terms of behavioral, neuropsychological, and neurophysiological data. Color is usually defined as psychophysical color or as perceived color. Behavioral data on categorical color perception, absolute judgement of colors, color coding, visual search, and visual awareness refer to the more cognitive aspects of color. These are of major importance in visual synthesis and spatial organization, as already shown by the Gestalt psychologists. Neuropsychological and neurophysiological findings provide evidence for an interrelation between cognitive color and spatial organization. Color also enhances planning strategies, as has been shown by studies on color and eye movements. Memory colors and the color- language connections in the brain also belong among the cognitive aspects of color.

This paper describes a new method for assisting the coloring work to achieve good color harmony in documents that contain color diagrams, figures, or images. Generally, good quality for a color document is achieved by color-element unification. Tone, a color element, is often used to arrange color harmony in a color document. A tone model is proposed that describes the color harmonization concept. An approximation formula is developed to describe the tone model. It enables simple correspondence between color coordinates and their tones. Using the tone model, users can generate a color palette that consists of colors whose subjective impression is unified. This palette is provided for coloring work, Because each of these colors has a uniform element in common, a color document, which is painted by colors from one palette, achieves the color harmonization. The tone model also provides another kind of human interface that allows the users to intuitively select a color by using the adjectives that give us the impression of color variations. Emphasis on a diagram part by selecting a color with a different tone is easily visualized. Using the prototype system, the time for coloring work was significantly reduced.

M odern desktops and application development environments give the user the ability to interactively select colors, fonts and graphical objects, but do not offer the user guidance in making these selections. Consequently, the user is free to produce illegible, unaesthetic, and jarring results. For example, the user can select dark blue text on a black background, fuscia and scarlet windows in the same application, and button text which is truncated because it doesn't fit on the button geometry. With additional iterations through the design process, users can often converge on a reasonable selection of colors, fonts and graphical objects, but this process is often time-consuming and frustrating, and may not result in a particularly good set of selections. Furthermore, the users of these systems may not be particularly sophisticated in issues of legibility, color harmony, or graphical design. Furthermore, when these users have difficulties, they very often interpret these as system failings, resulting in service calls to manufacturers. In a recent example, a user called to complain that his computer was broken because there was no text under his icons. He had inadvertently selected white for both the screen background color and the icon text color. In 1993, we introduced a general architecture for adding guidance to interactive systems (Rogowitz and Treinish, 1993) which we have extended to the domain of user interface design (Rogowitz and Rabenhorst, 1993) in a system called CRAFT (Color Rule and Font Tool). In this approach, the choices for each operation (e.g., selecting background color, selecting text color) are constrained by perceptual and aesthetic "rules". In this architecture, all related operations are linked, which means that every time the user makes a selection, the impact of that choice is reflected in the selections for other operations. For example, if the user selects a color for a window background, that information is fed back to the operation for selecting text color, where legibility rules use this information to constrain the set of text colors to those which provide sufficient luminance contrast for good legibility. This network of linked, intelligent operations help guide the user through the complex design space of colors and fonts. In this paper, we enhance the CRAFT system by developing the relationship between the rules and the information which drives them, with a special focus on fonts and font/color interactions.

This paper describes an approach to introducing guidance into choosing color representations, based on 'metavisualizations' and their use within formalized models of the visualization process. Metavisualizations are designed to allow fully manual color selection processes in a manner that can incorporate intelligence progressively. Autonomous agents, designed for each metavisualization, make sensible decisions based on stored knowledge in the context of the underlying models. Color selection can thus increment from fully manual through to fully automated. Choices exercised by the autonomous agents can at any stage be manually overridden by direct interaction with the metavisualization. The approach also encourages implicit user learning. Application examples show the mapping of various data sets into the perceptual gamut of a color display device, where metavisualizations provide an interactive graphical interface for controlling the choice of mappings.

This investigation is built upon our previous experiments on the experiential effect of color combinations. Within this study, we found that the systematic adjustment of color combinations can affect the magnitude of expressive response from individuals. We postulate that within each isobar of magnitude--low, medium, and high--there are distinct qualities of expression evoked, varying from stressful to harmonic. We designed an experimental procedure to measure responses by specifying various kinds of hue-to-hue alignment and controlling the variation of value and chroma contrast. Our experimental procedure and results are discussed in this paper.

Transmission bandwidth and memory, even with the trend of increased availability and lower costs, are resources always in demand and their use needs to be opthnized. From all the elements of information in digital form, images are the top consumers of memory and bandwidth. Therefore, image compression is becoming a tool of general use for storage and transmission purposes, bringing along conspicuous image degradation. Visually perceived degradation needs to be quantified in order to make appropriate decisions involving compression tradeoffs for different imaging applications. The ViDEOS (Video Display Engineering and Optimization System) project is a software system under development. It is a computational tool for engineering color LCD displays under different sets of specifications and evaluating performance under different conditions. Multiple levels of a design can be analyzed with ViDEOS without having to build expensive prototypes; from the electro-optical performance of specific crystal interfaces to image rendering of a complete design under different settings of trade-off parameters. One of the capabilities of the VIDEOS system is the prediction of image visibility when viewed on any display represented within the system. This capability is implemented by means of a model of human vision which is part of the VIDEOS system. One of the many interesting applications is evaluating the visibility of image artifacts like those produced by image compression. Our ongoing efforts reported here include producing images at different degrees of compression and then comparing decompressed versions with respect to the original image using the vision model.

Over the past several years there have been many attempts to incorporate perceptual masking models into image compression systems. Unfortunately, there is little or no information on how these models perform in comparison to each other. The purpose of this paper is to examine how two different perceptual models perform when utilized in the came coding system. The models investigated are the Johnson-Safranek and the Watson. They both develop a contrast masking threshold for DCT based coders. The coder used for comparison is the Baseline Sequential mode of JPEG. Each model was implemented and used to generate image dependent masking thresholds for each 8x8 pixel block in the image. These thresholds were used to zero out perceptually irrelevant coefficients, while the remaining coefficients were quantized using a perceptually optimal quantization matrix. Both objective and subjective performance data was gathered. Bit rate saving versus standard JPEG was computed, and a subjective comparison of images encoded with both models and the nonperceptual JPEG was run. The perceptually based coders gave greater compression with no loss in subjective image quality.

Self-similarities, commonly explored in fractal image compression, are usually translated into matches between two pools, the range and the domain blocks, which are different partitions of the same image to be encoded. Simple transformations on the domain blocks are used in order to obtain a better match. A root-mean-square error measure between a range block and a transformed domain block is used in the encoding process to quantify the performance of the matching process and subsequently the quality of the encoded imaged. Alternative measures can be used in fractal image compression to account for human visual perception. Simple strategies, such as block intensity weighting and block texture weighting, reduce perceptual degradation with only very little added computational cost. Weighted error measures in frequency domain, though computationally much more expensive, can provide a more natural model of visual perception such as the direct account for the contrast sensitivity function. A multiscale approach to encode the image details at different resolution is proposed to not only speed up the matching process between the range and the domain blocks but also provide a mechanism for multiscale representation.

We have developed new neural paradigms and simulators based on cortical models of the visual cortex in the brain. These simulators are currently developed for ultrahigh image compression in the MPEG4 research environment. The first significant results have just been achieved in very low bit rate coding (at 30 Kbits/s and less), with more acceptable aspects that fit with visual system requirement. Moreover neural tools seem to be generic and well adapted to the various input styles of TV sequences, thanks to their learning capabilities and architecture dedicated to all kinds of 2D and 3D analysis-synthesis processing.

We examined the sensitivity of the human visual system to video coding artifacts following a scene cut. Based on classical studies of visual masking, we would expect that sensitivity to coding artifacts would be reduced immediately following a cut. This study focused on the visibility of coding artifacts introduced by an MPEG-2 codec. Three image sequences were selected, each with a scene cut in the middle, such that image content in the first half was different from that in the second half. Using psychophysical methods, visibility threshold were estimated for coding artifacts inserted in the first, second, or third frame following the scene cut. Visibility thresholds were also estimated for coding artifacts that were inserted simultaneously in the first and second frames following the cut. it was found that visual masking effects were present in the first frame following a scene cut, but were virtually absent by the second or the third frame. The results of this research may be used to optimize the behavior of video coding algorithms at scene cuts.

In videoconferencing applications the perceived quality of the video signal is affected by the presence of an audio signal (speech). To achieve high compression rates, video coders must compromise image quality in terms of spatial resolution, grayscale resolution, and frame rate, and may introduce various kinds of artifact.s We consider tradeoffs in grayscale resolution and frame rate, and use subjective evaluations to assess the perceived quality of the video signal in the presence of speech. In particular we explore the importance of lip synchronization. In our experiment we used an original grayscale sequence at QCIF resolution, 30 frames/second, and 256 gray levels. We compared the 256-level sequence at different frame rates with a two-level version of the sequence at 30 frames/sec. The viewing distance was 20 image heights, or roughly two feet from an SGI workstation. We used uncoded speech. To obtain the two-level sequence we used an adaptive clustering algorithm for segmentation of video sequences. The binary sketches it creates move smoothly and preserve the main characteristics of the face, so that it is easily recognizable. More importantly, the rendering of lip and eye movements is very accurate. The test results indicate that when the frame rate of the full grayscale sequence is low (less than 5 frames/sec), most observers prefer the two-level sequence.

After some principal considerations about noise and the psychometric function,we present a model for the masking of a spatial frequency patten by nonwhite noise. From threshold elevation measurements by Stromeyer and Julesz, we derive a distribution function for the disturbing effect of one frequency on the detection of another frequency. We apply the resulting formulas on contrast sensitivity measurements with nonwhite noise by van Meeteren and Valeton. Next we use the same principles to derive a simple model for contrast discrimination. We consider the pedestal modulation as a nonwhite noise source of a single spatial frequency that hampers contrast detection. We show that contrast discrimination plots for different spatial frequencies will coincide if they are plotted in a normalized way. After a slight practical correction the model appears to be in very good agreement with measurements of various authors published in literature.

In a series of papers presented in 1994 at SPIE1' SID2 and IDRC3, we examined the grayscale/resolution trade-off for natural images displayed on devices with discrete pixellation, such as AMLCDs. In the present paper we extend this study by examining the grayscale/resolution trade-off for text images on discrete-pixel displays. Haiftoning in printing is an example of the grayscale/resolution trade-off. In printing, spatial resolution is sacrificed to produce grayscale. Another example of this trade-off is the inherent low-pass spatial filter of a CRT, caused by the point-spread function of the electron beam in the phosphor layer. On a CRT, sharp image edges are blurred by this inherent low-pass filtering, and the block noise created by spatial quantization is greatly reduced. A third example of this trade-off is text anti-aliasing, where grayscale is used to improve letter shape, size and location when rendered at a low spatial resolution. There are additional implications for display system design from the grayscale/resolution trade-off. For example, reduced grayscale can reduce system costs by requiring less complexity in the framestore, allowing the use of lower cost drivers, potentially increasing data transfer rates in the image subsystem, and simplifying the manufacturing processes that are used to construct the active matrix for AMLCD (active-matrix liquid-crystal display) or AMTFEL (active-matrix thin-film electroluminescent) devices. Therefore, the study of these trade-offs is important for display designers and manufacturing and systems engineers who wish to create the highest performance, lowest cost device possible. Our strategy for investigating this trade-off is to generate a set of simple test images, manipulate grayscale and resolution, predict discrimination performance using the ViDEOSSarnoff Human Vision Model (see Larimer et al., 1994), conduct an empirical study of thscrimination using psychophysical procedures, and verify the computational results using the psychophysical results.

Standard 1D gamma-correcting lookup tables do not compensate for adjacent pixel spatial nonlinearities along the direction of the display raster. These nonlinearities can alter the local mean luminance and contrast of the displayed image. Five steps are described for generating a 2D lookup table (LUT) that compensates for the nonlinearity. By adjusting the 2D LUT so it takes into account the inherent blur at light to dark transitions of the display system, the usable luminance range of the LUT can be extended while reducing the ringing artifact associated with luminance compensation. Use of the blur-compensated 2D LUT incurs no additional computational effort over an uncompensated 2D LUT. Matlab programs are included that can be used to generate a 2D LUT for a user's particular display system.

We propose a new operator for creating rotated dither threshold arrays. This new discrete one- to-one rotation operator is briefly explained. We analyze its application to different dispersed- dot dither arrays such as hexagonal dispersed dither arrays and 3X3 matrix-based Bayer- expanded dither arrays and compare the results with the ones obtained by rotating standard Bayer dither arrays. We show that the rotation operator introduced new lower-frequency components that, for example in the case of rotated dispersed-dot Bayer dither, produces a slight clustering effect, improving the tone reproduction behavior of the halftone patterns. In other cases, such as hexagonal dispersed dither, these new lower frequency components are responsible for strong interferences in the rotated halftone array. When applied to 3x3 matrix- based Bayer-expanded dither arrays, the rotation operator induces sequences of short horizontal and vertical patterns that have very good tone reproduction behavior in the dark tones. Besides their used in black and white printing, rotated dispersed-dot dither halftoning techniques have also been successfully applied to in-phase color reproduction on inkjet printers.

A new halftoning algorithm is proposed that incorporates the concept of delayed decision in error diffusion. This algorithm is similar to a recently proposed tree coding halftoner in that both examine at each pixel location a neighborhood of future (with respect to the scanning direction) pixels before deciding the value of the current output pixel. Because of the feedback in the error diffusion process, one can reduce the size of the 'look ahead' window compared to a typical tree-coding halftoner, resulting in a relatively low complexity. It is shown that this algorithm improves the quality of halftone images compared to those generated by traditional error diffusion.

Stochastic halftone screens can be used in ordered dither halftoning algorithms to generate binary image textures that have shaped power spectra. This paper will examine some techniques for modeling perception of binary texture. Alternative methods for constructing grayscale dot profiles will be discussed, with the aim of improving the visual optimality of the entire grayscale dot profile.

Frequency modulated (FM) halftoning or 'stochastic screening,' has attracted a great deal of attention in the printing industry in recent years. It has several advantages over conventional halftoning. But one serious problem that arises in FM halftoning is dot gain. One approach to stochastic screening uses a specially constructed halftone screen, the blue noise mask (BNM), to produce an unstructured and visually appealing pattern of halftone dots at any gray level. In this paper, we will present methods to correct dot gain with the BNM. Dot gain is related to the area-to-perimeter ration of printed spots. We can exploit this feature in different ways. At a medium level, a B>NM pattern will have 'connected' as well as 'isolated' dots. Normally, as we build down BNM patterns to lower levels, a specific number of white dots will be replace by black dots. Since connected white dots are more likely to be picked than isolated white dots, this will results in substantial dot gain because of the increasing number of isolated white dots. We show that it is possible to constrain the process of constructing a BNM such that isolated dots are preferentially removes, thus significantly reducing dot gain in a BNM.

Halftoning using a mask presents the advantage of computational simplicity and implementation speed. Blue-noise masks in particular have the additional advantage of producing visually pleasing halftones without periodic artifacts. However, halftones produced with a blue-noise mask are usually grainy and less sharp compared with error diffusion images. In this paper, we present an algorithm for the construction of a 'max-energy' mask that reduces the perceived graininess in the output halftone by controlling the local spectral energy in the dot profiles of the mask, i.e., the binary patterns that results after the mask is thresholded at a constant gray level. Dot profiles of the max-energy mask are shown and an example halftone is compared against a blue-noise mask halftone.

We reported in 1993 that luminance variations could distort the depicted scene and influence the accuracy of judgment of spatial layout in computer-generated imagery. Since that time, there have been reports that suggest that it is luminance contrast, rather than luminance per se, that functions as a depth cue (based on the argument that a negative luminance-contrast gradient simulates the natural world effect of atmospheric perspective). Since a negative luminance gradient can also simulate atmospheric perspective, and neither a negative luminance gradient or a negative luminance-contrast gradient can be accurately represented with real-tim computer-generated images, we have considered the influence of each of these gradients on observers' perception of scene layouts generated by computer, in order to assess their input on the perceptions of depicted layout. Our findings suggest that both luminance gradients and luminance contrast gradients can affect judgments of spatial layout. However, the nature of their influence depends on the amount and type of other depth information available on the screen. The results are consistent with a model of depth cue combination that weights each cue in relation to the other available cues, with implications for the resolution of conflicting cues. In addition, we suggest that these is a need for convention for luminance contrast measurement of computer monitors, to enable closer comparison between research papers.

The quality of 3D images is usually perceived as being better than that of 2D images. In 2D images, edge sharpness is responsible for the quality. Accordingly, we examined the effects of the stereoscopic perception of edge sharpness on image quality by employing the categorized pair comparison method. We compared the sharpness of two types of natural 3D images with 2D images and found that the sharpness of 3D was not always perceived as being sharper than 2D. Next experiments were carried out to examine the role of rivalry on edge sharpness using a geometric image to which a local rivalry was artificially added. Results showed that the artificial local rivalry reduced the edge sharpness. Further experiments compared the sharpness of the vertical or horizontal edges in 3D images, which consisted of two planes (near and far), with that in 2D images. Consequently, the 3D vertical edge was as sharp as the 2D one, while the horizontal edge was perceived as being sharper than the 2D. As a result, the improvement in edge sharpness raises overall 3D image quality. Moreover, the perception of an occluded region in 3D images affected the edge sharpness. We also refer to the relationship between the perception and rivalry that might have caused interocular suppression.

The induced motion from visual information is discussed in showing the stable and moving images with or without disparities. Experiments were conducted to measure body sway, eye movement, and head movement under plane depth visual stimuli, to clarify the effect caused by the depth information for induced motion. Several obvious points emerged from the analysis of the experimental data. The minimal visual area needed to make the body stale is about 45 degrees for the plane stimuli, and is about 22.5 degrees of 45 degrees for the depth stimuli. It was also found that there are large differences in body sway between the plane stimuli and the depth stimuli after target moving. For the moving visual stimuli without disparity, the body sway was found to depend on the visual angle, but in the other case, it was found to be independent of the visual angle. This is due to the difference in OKN values from the measurement of eye movement. The frequency of emergence and the amplitude of OKN in the depth stimuli were larger than those of OKN in the plane stimuli. It was additionally observed that the low-frequency components of body sway were equal to those of head movement in these experiments.

This paper reports on the visibility of spatial details located in those regions of a stereo pair that are seen by one eye only. In normal viewing, these regions correspond to areas of a scene that are occluded by nearer objects for one eye or the other--so-called half occluded regions (or monocular regions). We assumed that the resolving power of human vision in binocular regions would be higher than that in half-occluded regions because of the effect of binocular summation. If it were true, high spatial frequency components in half-occluded regions could be reduced without impairment of the subjective image quality. A comparison test of the visual resolving power for (spatially) low-pass filtered and nonfiltered patterns in half-occluded regions was carried out. The experiment showed, however, that any impairment visible in binocular regions was also visible in half-occluded regions. Thus the assumption has been falsified, and for the purpose of stereo image coding no irrelevant data elimination can be carried out in half-occluded regions.

Visual illusions abound in normal vision--illusions of clarity and completeness, of continuity in time and space, of presence and vivacity--and are part and parcel of the visual world inwhich we live. These illusions are discussed in terms of the human visual system, with its high- resolution fovea, moved from point to point in the visual scene by rapid saccadic eye movements (EMs). This sampling of visual information is supplemented by a low-resolution, wide peripheral field of view, especially sensitive to motion. Cognitive-spatial models controlling perception, imagery, and 'seeing,' also control the EMs that shift the fovea in the Scanpath mode. These illusions provide for presence, the sense off being within an environment. They equally well lead to 'Telepresence,' the sense of being within a virtual display, especially if the operator is intensely interacting within an eye-hand and head-eye human-machine interface that provides for congruent visual and motor frames of reference. Interaction, immersion, and interest compel telepresence; intuitive functioning and engineered information flows can optimize human adaptation to the artificial new world of virtual reality, as virtual reality expands into entertainment, simulation, telerobotics, and scientific visualization and other professional work.

The determinants of visual scale (size and distance) under monocular viewing are still largely unknown. The problem of visual scale under monocular viewing becomes readily apparent when one moves about within a virtual environment. It might be thought that the absolute motion parallax of stationary objects (both in real and virtual environments), under the assumption of the stationarity, would immediately determine their apparent size and distance for an observer who is walking about. We sought to assess the effectiveness of observer- produced motion parallax in scaling apparent size and distance within near space. We had subjects judge the apparent size and distance of real and virtual objects under closely matched conditions Real and virtual targets were 4 spheres seen in darkness at eye level. the targets ranges in diameter from 3.7 cm to 14.8 cm and were viewed monocularly from difference distances, with a subset of the size/distance combinations resulting in projectively equivalent stimuli at the viewing origin. subjects moved laterally plus and minus 1 m to produce large amounts of motion parallax. when angular size was held constant and motion parallax acted as a differential cue to target size and distance, judged size varied by a facto or 1.67 and 1.18 for the real and virtual environments, respectively, well short of the fourfold change in distal size. Similarly, distance judgments varied by factors of only 1.74 and 1.07 respectively. We conclude that absolute motion parallax only weakly determines the visual scale of nearby objects varying over a fourfold range in size.

Segmentation is one of the most important image processing steps toward pattern recognition and image understanding and a significant step toward image compression and coding. In this paper, a fuzzy technique is presented that resolves several difficult issues related to image segmentation. The largest area of an image can be easily segmented and the association of most of the pixels to different segments determined from the edge information. It is, however, the pixels between segments and in high frequency areas whose relationship to adjacent segments is not so clear. These pixels can belong to the matte, highlight, or shadow area of the closest segment, or in association with neighboring pixels make up a smaller segment. the dichromatic reflection model is used to merge highlight and shadow areas with the matte ones. The approximate or fuzzy reasoning approach proposed here provides a more humanlike segmentation of images.

The computer simulation model for adaptive segmentation of 2D visual objects is based on several psychological and neurophysiological principles. the following stages of visual information processing are simulated: (1) the preliminary image estimation, which is made by the analysis of the image intensity distribution; (2) they synthesis of the control process for primary image segmentation based on the results of initial image estimation; (3) the invocation of 'sustained' and 'transient' mechanisms of image description; the first one extracts information about the image contours, while the second selects the fragments with homogeneous intensity; (4) the evaluation of primary image descriptions based on the analysis of image spatial characteristics; (5) the synthesis of a processing scheme that includes the assignment and the adjustment of processing operations for each image fragment, produced from the previous analysis; (6) the final image segmentation stage, carried out by the processing of 'sustained' and 'transient' image descriptions. The computer program for adaptive image segmentation has been tested on various 2D grayscale images.

This paper reports results from interdisciplinary research: a framework for the integration of multiple information in image analysis called 'information fusion in image understanding' is applied to provide a new visualization scheme for diagnosis and treatment of the human retina. This framework deals with representations and processes at all levels of abstraction. It is used to represent anatomical and pathological knowledge, to extract significant features from the input channels, and to obtain a complex diagnosis by means of fusion. Each patient is examined in six steps using a scanning laser ophthalmoscope (SLO) providing several spectral channels and aperture settings, as well as static scotometry to measure scotoma (areas with a loss of visual function). Feature extraction processes yield dark (fovea) and bright (leakage) blobs at several scales,clusters of scotoma measures, tubes (blood vessels), and circular areas (optic disc) in six different image description. By affine matching, the fusion system establishes spatial relationships between the features of each image description. Spatial reasoning using these relationships leads to a fusion of symbolic information without the necessity or prior manual registration of the input channels. Finally, the extracted features are arranged in different overlays to visualize a 'map of anatomical features and pathological changes.' SLO examinations using the above scheme were carried out with more than 50 patients suffering from age-related macular degeneration at the Vienna Eye Clinic. The application of information fusion has now lead to a new concept about the nature of the disease, to new diagnostic capabilities, and to a clear distinction between three different classes of therapy.

This paper considers shape constancy, which is a fundamental perceptual phenomenon. Shape constancy refers to the fact that the percept of the shape of a given object remains constant despite changes in the object's retinal image. The image may change because of changes in the orientation of the object relative to the observer. In conventional approaches based on the concept of groups of transformations, formulating a theory of shape constancy requires that a group that adequately represents the viewing conditions be identified first. As a results, a given theory holds only for the assumed group, and therefore many theories are needed, one theory for each viewing condition. It is conjectured that explaining shape constancy requires a new approach, in which one theory holds across different viewing conditions. In this new approach, shape constancy is explained by using geometrical properties of image formation under perspective. Image formation is not represented by a group of transformations, but it leads to a theory that is more general than prior theories it explains shape constancy in the case of monocular and binocular viewing for both the stationary and the active observer.

Image sequences can now be synthesized by compositing elements--objects, backgrounds, people--from preexisting source images. What are the rules that govern the placement of these elements in their new frames? If geometric projection is assumed to govern the composition, there is some freedom to move the elements about, but the range of placements is quite limited. Furthermore, projective geometry is not perceptually ideal--think of the distortions seen at the edges of wide-angle photographs. These distortions are not found in perspective paintings: painters modify the portrayal of certain objects to make them perceptually correct. The paper first reviews projective geometry in the context of image compositing and then introduces perceptually based composition. Its basis in human vision in analyzed and its application to image compositing discussed.

Object detection involves looking for one of a large set of object subimages in a large set of background images. Image discrimination models predict the probability that an observer will detect a difference between two images. We find that discrimination models can predict the relative detectability of objects in different images, suggesting that these simpler models may be useful in some object detection applications. Six images of a vehicle in an otherwise natural setting were altered to remove the vehicle and mixed with the original image in various proportions. Nineteen observers rated the 24 images for the presence of a vehicle. The pattern of observer detectabilities for the different images was predicted by three discrimination models. A Cortex transform discrimination model, a contrast sensitivity function filter model, and a root-mean-square difference predictor based on the digital image values gave prediction errors of 15%, 49%, and 46%, respectively. Two observers given the same images repeatedly to make the task a discrimination task rated the images similarly, but had detectabilities a factor of two higher.

The relation between perceptual image quality and naturalness was investigated by varying the colorfulness and hue of color images of natural scenes. These variations were created by digitizing the images, subsequently determining their color point distributions in the CIELUV color space and finally multiplying either the chroma value of the hue angle of each pixel by a constant. During the chroma/hue-angle transformation the lightness and hue-angle/chroma value of each pixel were kept constant. Ten subjects rated quality and naturalness on numerical scales. The results show that both quality and naturalness deteriorate as soon as hues start to deviate from the ones in the original image. Chroma variation affected the impression of quality and naturalness to a lesser extent than did hue variation. In general, a linear relation was found between image quality and naturalness. For chroma variation, however, a small but systematic deviation could be observed. This deviation reflects the subjects' preference for more colorful but, at the same time, somewhat unnatural images.

Gaze-contingent interfaces must provide adequate peripheral information to the viewer to preserve perceptual fidelity. specifically, it has been shown that preattentive (or preview) benefit must be preserved so that scene perception is not disrupted. In this paper we review recent attempts at peripheral degradation of digital imagery. We suggest that to be successful such degradation must preserve potential future visual attractors and, furthermore, not introduce artificial ones. We present a simple multiresolution image processing approach that can be utilized for this purpose. The feasibility of gaze-contingent processing has recently been questioned. In their paper, Stelmach and Tam processed images by low-pass filtering, effectively smoothing extrafoveal regions. The authors then quantized DCT coefficients in the periphery, introducing blocking artifacts. In our paper, we simulate these results and claim that neither of these methods is suitable for GC interfaces. Alternatively, we implement a simple multiple region of interest (ROI) multiresolution scheme in an attempt to degrade the periphery while preserving attentional cues. We evaluate three variants of this approach: a linear degradation function, a nonlinear function, and a function matching human visual system (HVS) acuity. The HVS-matching multiple-ROI algorithm gives good compression and alleviates the degradation of potential visual attractors. Furthermore, MIP mapping offers efficient implementation of the algorithm, making it a good candidate for gaze-contingent applications.