Beginning with this issue, Jan Allebach of Purdue University becomes the new Editor of JEI. I have been Editor for six years and it is time for someone to bring new energy and new ideas to the editorship.

This collection of papers commemorates the tenth anniversary of the IS&T/ SPIE Conference on Human Vision and Electronic Imaging. These papers represent major trends in the conference and demonstrate the interplay between real-world imaging applications and vision research.

The field of electronic imaging has made incredible strides over the past decade producing systems with higher signal quality, complex data formats, sophisticated operations for analyzing and visualizing information, advanced interfaces, and richer image environments. Since electronic imaging systems and applications are designed for human users, the success of these systems depends on the degree to which they match the features of human vision and cognition. This paper reviews the interplay between human vision and electronic imaging, describing how the methods, models and experiments in human vision have influenced the development of imaging systems, and how imaging technologies and applications have raised new research questions for the vision community. Using the past decade of papers from the IS&T/SPIE Conference on Human Vision and Electronic Imaging as a lens, we trace a path up the ‘‘perceptual food chain,’’ showing how research in low-level vision has influenced image quality metrics, image compression algorithms, rendering techniques and display design, how research in attention and pattern recognition have influenced the development of image analysis, visualization, and digital libraries systems, and how research in higher-level functions is involved in the design of emotional, aesthetic, and virtual systems.

The growth of digital video has given rise to a need for computational methods for evaluating the visual quality of digital video. We have developed a new digital video quality metric, which we call DVQ (digital video quality) [A. B. Watson, in Human Vision, Visual Processing, and Digital Display VIII, Proc. SPIE 3299, 139– 147 (1998)]. Here, we provide a brief description of the metric, and give a preliminary report on its performance. DVQ accepts a pair of digital video sequences, and computes a measure of the magnitude of the visible difference between them. The metric is based on the discrete cosine transform. It incorporates aspects of early visual processing, including light adaptation, luminance, and chromatic channels; spatial and temporal filtering; spatial frequency channels; contrast masking; and probability summation. It also includes primitive dynamics of light adaptation and contrast masking. We have applied the metric to digital video sequences corrupted by various typical compression artifacts, and compared the results to quality ratings made by human observers.

Sensitivity and resolution reduction as a function of eccentricity account for one of the largest sources of compression in vision. However, utilization of this visual property has been limited to systems that directly measure the viewer’s gaze position. We have applied visual eccentricity models to videophone compression applications without using eye tracking by combining the visual model with a face tracking algorithm. In lieu of a gaze detector, we assume the gaze will be directed to the faces appearing in images. The incorporation of resolution as well as sensitivity-based eccentricity models in a low bit rate video-compression standard (H.263) will be discussed. For videophone applications, we achieve up to a 50% reduction in bit rate while retaining similar image quality. Problems arising from the increased temporal sensitivity of the periphery, despite its reduced spatial bandwidth, are also discussed.

Designers of imaging systems, image processing algorithms, etc., usually take for granted that methods for assessing perceived image quality produce unbiased estimates of the viewers’ quality impression. Quality judgments, however, are affected by the judgment strategies induced by the experimental procedures. In this paper the results of two experiments are presented illustrating the influence judgment strategies can have on quality judgments. The first experiment concerns contextual effects due to the composition of the stimulus set. Subjects assessed the sharpness of two differently composed sets of blurred versions of one static image. The sharpness judgments for the blurred images present in both stimulus sets were found to be dependent on the composition of the set as well as the scaling technique employed. In the second experiment subjects assessed either the overall quality or the overall impairment of manipulated and standard JPEG-coded images containing two main artifacts. The results indicate a systematic difference between the quality and impairment judgments that could be interpreted as instruction-based different weighting of the two artifacts. Again, some influence of scaling technique was observed. The results of both experiments underscore the important role judgment strategies play in the psychophysical evaluation of image quality. Ignoring this influence on quality judgments may lead to invalid conclusions about the viewers’ impression of image quality.

The classical approach in vision research—the derivation of basically linear filter models from experiments with simple artificial test stimuli—is currently undergoing a major revision. Instead of trying to keep the dirty environment out of our clean labs we put it now right into the focus of scientific exploration. An increasing number of scientists are using natural images in their experimental work, and concepts from statistics and information theory are employed for the theoretical modeling of the results. The new approach has a close relation to basic engineering strategies for electronic image processing, since its major concept is that biological sensory systems exploit the statistical redundancies of the environment by appropriate neural transformations. The standard engineering methods are not sufficient, however. Even such a basic biological feature as orientation selectivity requires the consideration of higher-order statistics, like multivariate wavelet statistics, cumulants, or polyspectra. Furthermore, there exists an abundance of nonlinear phenomena in biological vision, for example the phase invariance of complex cells, cortical gain control, end-stopping, and a variety of extra-classical receptive field properties. These amount to nonlinear combinations of linear wavelet filter outputs, which are required to exploit higherorder statistical dependencies, and make it necessary to consider unconventional modeling approaches like differential geometry or Volterra–Wiener systems. By use of such methods we cannot only gain a deeper understanding of the adaptation of the visual system to the complex natural environment, but we can also make the biological system an inspiring source for the design of novel strategies in electronic image processing.

We examined judgments of distortions in images of faces and how they are influenced by visual adaptation to distorted images. Two-dimensional image arrays were generated by varying the degree of local expansion or contraction along the horizontal or vertical axis of the image. Images along different directions within the array were varied in staircases while observers rated the images as either ‘‘normal’’ or ‘‘distorted.’’ The reversal points for the staircases define the gamut of distortions that observers accept as a normal variant of the original image, and they are well described by ellipses. Under neutral adaptation this gamut is roughly centered on the original image. However, after viewing a distorted face, the images perceived as normal are strongly biased toward the adapting distortion, and the range of the gamut increases along the axis of the distortion. Thus changes in the state of adaptation can markedly alter the perception of images, by altering both the perceived neutral point and the sensitivity to variations in images.

The experiments in this paper demonstrate that ‘‘Early Vision’’ mechanisms can account for the appearance of three ‘‘Diamond Wall’’ demonstrations, without reliance on apparent illumination, transparency, apparent depth, and junctions. The first set (Section 2) demonstrates that simultaneous contrast and the Katz– Albers effect can explain the appearance of a Diamond Wall display. The second set (Section 3) reviews the ‘‘Straight Edge’’ experiments designed to show changes in lightnesses consistent with perceived illumination. These experiments showed that very-low-spatialfrequency sampling causes both these effects, and White’s effect. The third set (Section 4) applies complex Early Vision models to images associated with the ‘‘High Vision’’ lightness hypotheses. The argument is that flat displays, which are perceived as flat, require a quite complex visual mechanism just to account for the properties of flat lightnesses. Any experimental verification of the existence of High Vision lightness mechanisms should be tested first with realistic complex Early Vision models. The results show that Early Vision mechanisms can account for appearances in Diamond Wall experiments. If Early Vision mechanisms can explain these results, then these experiments cannot be used as evidence for the existence of High Vision mechanisms.

The repetitive scanpath eye movement, EM, sequence enabled an approach to the representation of visual images in the human brain. We supposed that there were several levels of binding—semantic or symbolic binding; structural binding for the spatial locations of the regions-of-interest; and sequential binding for the dynamic execution program that yields the sequence of EMs. The scanpath sequences enable experimental evaluation of these various bindings that appear to play independent roles and are likely located in different parts of the modular cortex. EMs play an essential role in top-down control of the flow of visual information. The scanpath theory proposes that an internal spatial-cognitive model controls perception and the active looking EMs. Evidence supporting the scanpath theory includes experiments with ambiguous figures, visual imagery, and dynamic scenes. It is further explicated in a top-down computer vision tracking scheme for telerobots using design elements from the scanpath procedures. We also introduce procedures—calibration of EMs, identification of regions-of-interest, and analysis and comparison programs—for studying scanpaths. Although philosophers have long speculated that we see in our mind’s eye, yet until the scanpath theory, no strong scientific evidence was available to support these conjectures.

The perception of an image by a human observer is usually modeled as a parallel process in which all parts of the image are treated more or less equivalently, but in reality the analysis of scenes is a highly selective procedure, in which only a small subset of image locations is processed by the precise and efficient neural machinery of foveal vision. To understand the principles behind this selection of the ‘‘informative’’ regions of images, we have developed a hybrid system that consists of a combination of a knowledgebased reasoning system with a low-level preprocessing by linear and nonlinear neural operators. This hybrid system is intended as a first step towards a complete model of the sensorimotor system of saccadic scene analysis. In the analysis of a scene, the system calculates in each step which eye movement has to be made to reach a maximum of information about the scene. The possible information gain is calculated by means of a parallel strategy which is suitable for adaptive reasoning. The output of the system is a fixation sequence, and finally, a hypothesis about the scene.

Bottom-up or saliency-based visual attention allows primates to detect nonspecific conspicuous targets in cluttered scenes. A classical metaphor, derived from electrophysiological and psychophysical studies, describes attention as a rapidly shiftable ‘‘spotlight.’’ We use a model that reproduces the attentional scan paths of this spotlight. Simple multi-scale ‘‘feature maps’’ detect local spatial discontinuities in intensity, color, and orientation, and are combined into a unique ‘‘master’’ or ‘‘saliency’’ map. The saliency map is sequentially scanned, in order of decreasing saliency, by the focus of attention. We here study the problem of combining feature maps, from different visual modalities (such as color and orientation), into a unique saliency map. Four combination strategies are compared using three databases of natural color images: (1) Simple normalized summation, (2) linear combination with learned weights, (3) global nonlinear normalization followed by summation, and (4) local nonlinear competition between salient locations followed by summation. Performance was measured as the number of false detections before the most salient target was found. Strategy (1) always yielded poorest performance and (2) best performance, with a threefold to eightfold improvement in time to find a salient target. However, (2) yielded specialized systems with poor generalization. Interestingly, strategy (4) and its simplified, computationally efficient approximation (3) yielded significantly better performance than (1), with up to fourfold improvement, while preserving generality.

Psychophysical experiments were conducted on PicHunter, a content-based image retrieval (CBIR) experimental prototype with the following properties: (1) Based on a model of how users respond, it uses Bayes’s rule to predict what target users want, given their actions. (2) It possesses an extremely simple user interface. (3) It employs an entropy-based scheme to improve convergence. (4) It introduces a paradigm for assessing the performance of CBIR systems. Experiments 1–3 studied human judgment of image similarity to obtain data for the model. Experiment 4 studied the importance of using: (a) semantic information, (b) memory of earlier input, and (c) relative and absolute judgments of similarity. Experiment 5 tested an approach that we propose for comparing performances of CBIR systems objectively. Finally, experiment 6 evaluated the most informative display-updating scheme that is based on entropy minimization, and confirmed earlier simulation results. These experiments represent one of the first attempts to quantify CBIR performance based on psychophysical studies, and they provide valuable data for improving CBIR algorithms. Even though they were designed with PicHunter in mind, their results can be applied to any CBIR system and, more generally, to any system that involves judgment of image similarity by humans.

Three-dimensional computer graphic displays contain the information that viewers use to understand the shapes of depicted objects. We have devised techniques for carrying out experiments that examine the factors that influence the accuracy of this shape understanding. We construct simple convex objects using the superquadric equation. The objects are graphically rendered and displayed in rotation in a variety of perceptual tasks. In this paper, we summarize the results of our previous experiments that examined the relative contribution of shading and occluding contour, and the effects of including specular reflectance. We also report experiments designed to determine the influence that light source direction and surface striping have on shape understanding.

Palette synthesis and analysis tools have been built based upon a model of color experience. This model adjusts formal compositional elements such as hue, value, chroma, and their contrasts, as well as size and proportion. Clothing and household product designers were given these tools to guide their selection of seasonal palettes in the production of the private-label merchandise in a large retail chain. The designers chose base palettes. Accents to these palettes were generated with and without the aid of the color tools. These palettes are compared by using perceptual metrics and interviews. The results are presented.

In this special section we have the opportunity to explore the science and technology of image processing and scene analysis as applied to the field of machine vision. Machine vision can be described as the integration of sensors and optics, computers, algorithms, and robotics to automate manufacturing inspection and characterization for quality and process control.

A vision system capable of imaging and detecting defects on reflective nonplanar surfaces in the production line at a high cadence is presented in this paper. Defects are typically dust located under the metallic layer of packaging products used in cosmetic industries. To realize this processing, structured lighting which reveals the defects in the image is proposed. Defects appear clearly in the images like a set of brilliant pixels in dark zones. The signature of the defect is then obtained. The size of this signature does not depend linearly on the size of the defect. It is a function of the observation angle. In order to realize a precise and robust process, the necessity of acquiring several images of the same defect is demonstrated. Because of the acquisition rate, it has been necessary to optimize image processing time by the means of an original laplacian filter and of high level techniques of programming. Results obtained using this detection system are finally presented.

This paper is a contribution to a benchmark concept definition and it proposes a set of image segmentation algorithms which should be included in any benchmark for vision system. The proposed alagorithms allow us to predict, estimate, and quantify many image/vision system parameters such as data acquisition rate, local and global system communications, machine capability to support recursive processings, data reorganization, different programming/ implementation models, system scalability, etc. The proposed three algorithms calculate: multithreshold image connect component labeling, image region identification (through two complementary steps: isotropic seed region growing and pixel fusion), and data reorganization. Formalized algorithms, which can be implemented on different plate-forms (sequential, parallel, distributed) supporting different programming and implementation paradigms (models) are given. Machine performance analysis and quantification via matching coefficients are proposed as well.

This article describes how under-resolved images of bar codes may be read by suitable processing. A two-dimensional image of a bar code with insufficient resolution to be able to resolve the individual bars is processed to give a high-resolution image. For this to work, the bar code (or camera) must be slightly rotated to give a fraction of a pixel offset between rows. Since the bars are straight, the offset relative to the first complete row of the bar code increases linearly with vertical position in the image. This offset between rows results in a shift in phase that is proportional to both offset and frequency. A phase image is created by Fourier transforming each row in the image, and retaining the phases. By subtracting the first row from subsequent rows of the phase image, a surface is fitted to give the offset between rows. A high-resolution image is then formed by interleaving the pixel values from rows where the offset is nearest to the new pixel spacing. This image appears blurred because of the area sampling caused by the sensor, combined with the low pass response of the camera electronics. By modeling the image capture system, the point spread function may be estimated and then removed by using inverse filtering in the frequency domain. The offset between the rows is then removed by using a linear phase filter. This allows the rows within the resultant image to be averaged to reduce noise.

Color histograms computed from the normalized and hue color space are negatively affected by sensor noise due to the instability of these color space transforms at many RGB values. To suppress the effect of sensor noise, in this paper density estimations are computed using variable kernels. To that end, models are proposed for the propagation of sensor noise through the normalized and hue colors. As a result, not only the hue and normalized color values are known, but also the associated uncertainty. This twofold information is used to derive the parameterization of the variable kernel used for the density estimation. It is empirically verified that the proposed method compares favorably to the traditional histogram.

The detection and classification of faults is a major task for optical nondestructive testing in industrial quality control. Interferometric fringes, obtained by real-time optical measurement methods, contain a large amount of image data with information about possible defect features. This mass of data must be reduced for further evaluation. One possible way is the filtering of these images by applying the adaptive wavelet transform, which has been proved to be a capable tool in the detection of structures with definite spatial resolution. In this paper we show the extraction and classification of disturbances in interferometric fringe patterns, the application of several wavelet functions with different parameters for the detection of faults, and the combination of wavelet filters for fault classification. Examples of fringe patterns of known and varying fault parameters are processed showing the trend of the extracted features in order to draw conclusions concerning the relation between the feature, the filter parameter, and the fault attributes. Real-time processing was achieved by importing video sequences in a hybrid optoelectronic system with digital image processing and an optical correlation module. The optical correlator system is based on liquidcrystal spatial light modulators, which are addressed with image and filter data. Results of digital simulation and optical realization are compared.

Dimensional control by artificial vision is becoming a standard tool for industrialists interested in such remote and without contact measurement methods. The expected accuracy of those systems is dependent on camera resolution. High precision requires very costly charge coupling device sensors and frame grabbers. The proposed method tends to increase significantly the precision of dimensional measurements without increasing the hardware complexity. This algorithm is also quite robust against noisy images as it can be encountered in real world imaging; a precision of 1/16 pixel can easily be obtained with signal to noise ratio=2 dB. Our approach aims at improving the edge detection process involved in dimensional control by artificial vision. A lot of edge detection techniques with pixel resolution are well known and some of them are designed in order to be robust against image corruption. On the other hand B-spline interpolation methods have been considerably improved and popularized by the signal processing techniques proposed by M. Unser et al. An algorithm resulting from the merging of these two ideas is proposed in this paper. In this algorithm, the interpolation is prepared by an optimized filtering and by a detection of local maxima of gradient.

This paper provides an overview of the state of the art of digital shearography. It covers techniques for speckle averaging, fringe pattern reconstruction, several novel phase modulators, high speed speckle pattern registration. Examples in experimental mechanics including measurements of strain, vibration, shape, and nondestructive testing are also provided.

We present a volumetric approach to three-dimensional (3D) object modeling that differs from previous techniques in that both object texture and geometry are considered in the reconstruction process. The motivation for the research is the simulation of a thermal tire inspection station. Integrating 3D geometry information with two-dimensional thermal images permits the thermal information to be displayed as a texture map on the tire structure, enhancing analysis capabilities. Additionally, constructing the tire geometry during the inspection process allows the tire to be examined for structural defects that might be missed if the thermal data were textured onto a predefined model. Experimental results demonstrate the efficacy of the proposed approach and quantitative experiments indicate that the volumetric integration technique compares favorably to a state-of-the-art, mesh-based integration approach in terms of geometrical accuracy. Future research goals are also noted.

We present a surface profiler that produces unambiguous, fringe-free, surface profiles. Each point in the output image has a modulus of the complex degree of coherence (|µxy|) which varies monotonically as the optical path difference is increased from zero. To calculate |µxy|, the profiler incorporates a novel Stokes polarimeter employing a polarization image splitter and ferroelectric liquid crystal wave plates. Rapid wave plate addressing, coupled with the image splitter, allows data to be collected in three video frames. We observe that manufacturing inaccuracies in wave plate retardance and alignment introduce an error to the calculated |µxy| which can reintroduce ambiguity into the height measurement. We present a numerical correction which allows recovery |µxy| from the measured data, thereby returning us to an unambiguous surface profile. Results for various surfaces, showing the effects of systematic error and applied correction terms, are given.

Many optical range sensors use an equal angle increment (EAI) sampling. Such an EAI model generates data for surface geometrical description that have to be converted, in many applications, into data which meet the desired equal distance increment orthographic projection model. In three-dimensional (3D) image processing, a 3D interpolation scheme is needed to resample the range data into spatially equal-distance sampling data. In this paper, a resampling approach using a B-spline surface fitting approach is proposed. The first step is to select a new scale for all X, Y, and Z directions. The size of the new range image and the new coordinates of each point are then computed according to the actual references of (x,y,z) coordinates and the new scale. The new range data are interpolated using a B-spline surface fitting approach. The experiments show that this 3D interpolation approach provides a geometrically accurate solution for many industrial applications which deploy the EAI sampling sensors.

We herein present a paper dedicated to machine vision systems used in the metallurgy industry for process control. The four different stages (lighting, acquisition, processing and decision) of a machine vision system are briefly recalled through various systems implanted in the steel industry. The main focus of the present paper is set on the control of surface treatment processes. Surface temperature measurements being often a crucial point in the process control, after having quickly presented noncontact temperature measurements, two machine vision systems relying on temperature measurements are carefully described. The two controlled processes: the laser cladding and the high frequency welding of metallic pipes, are presented, and results as well as limits of the two systems are shown.

Marine biologists traditionally determine the presence and quantities of different types of fish by dragging nets across the bottom, and examining their contents. This method, although accurate, kills the collected fish, damages their habitat, and consumes large quantities of resources. This paper presents an alternative, a machine vision system capable of determining the presence of fish species. Illumination presents a unique problem in this environment, and the design of an effective illumination system is discussed. The related issues of object orientation and measurement are also discussed and resolved. Capturing images of fish in murky water also presents challenges. An adaptive thresholding technique is required to appropriately segment the fish from the background in these images. Mode detection, and histogram analysis are useful tools in determining these localized thresholds. It is anticipated that this system, created in conjunction with the Rutgers Institute for Marine and Coastal Science, will effectively classify fish in the estuarine environment.

This paper presents an architecture for real-time generic convolution of a mask and an image. The architecture is intended for fast low-level image processing. The field programmable gate array (FPGA)-based architecture takes advantage of the availability of registers in FPGAs to implement an efficient and compact module to process the convolutions. The architecture is designed to minimize the number of accesses to the image memory and it is based on parallel modules with internal pipeline operation in order to improve its performance. The architecture is prototyped in a FPGA, but it can be implemented on dedicated very-large-scale-integrated devices to reach higher clock frequencies. Complexity issues, FPGA resources utilization, FPGA limitations, and real-time performance are discussed. Some results are presented and discussed.

The early work on separated-kernel image processing using finite state machines (SKIPSM) concentrated mainly on binary processing which provided a fast means to implement binary morphology with very large structuring elements in a single pass using relatively small lookup tables. The basic concept can be extended to certain gray-scale processing techniques with a useful improvement in processing speed. Gray-scale SKIPSM algorithms present more of a challenge because the extremely large number of states that exist make the use of the lookup tables for the direct implementation of finite state machines impractical. However, as will be demonstrated, it is still possible to realize significant speed advantages with SKIPSM-implemented gray-scale processing using appropriate coding techniques. This paper describes two such techniques, grayscale morphology and Gaussian low-pass filtering.

Nowadays, vision-based inspection systems are present in many stages of the industrial manufacturing process. Their versatility, which permits us to accommodate a broad range of inspection requirements, is, however, limited by the time consuming system setup performed at each production change. This work aims at providing a configuration assistant that helps to speed up this system setup, considering the peculiarities of industrial vision systems. The pursued principle, which is to maximize the discriminating power of the features involved in the inspection decision, leads to an optimization problem based on a high-dimensional objective function. Several objective functions based on various metrics are proposed, their optimization being performed with the help of various search heuristics such as genetic methods and simulated annealing methods. The experimental results obtained with an industrial inspection system are presented. They show the effectiveness of the presented approach, and validate the configuration assistant as well.

The organization of the hardware and software as well as the application of a remotely operated vision system are all described. The system enables a vision engineer, working a long way from the site of a proposed installation, to experiment with a prototyping image-acquisition subsystem that is already in situ. Once suitable images have been obtained, he then designs a suitable image processing algorithm. In this way, the vision engineer is able to construct a prototype vision system that can be used to test the algorithm in a realistic operating environment, collect performance statistics and possibly refine the algorithm in light of the experience gained. The objective of this research is to streamline the design and prototyping phases for online factory-floor inspection systems, so that the optimum benefit can be obtained from the efforts of a skilled vision engineer.

Quality of Service (QoS)-guarantee in real-time communication for multimedia applications is significantly important. An architectural framework for multimedia networks based on substreams or flows is effectively exploited for combining source and channel coding for multimedia data. But the existing frame by frame approach which includes Moving Pictures Expert Group (MPEG) cannot be neglected because it is a standard. In this paper, first, we designed an MPEG transcoder which converts an MPEG coded stream into variable rate packet sequences to be used for our joint source/channel coding (JSCC) scheme. Second, we designed a classification scheme to partition the packet stream into multiple substreams which have their own QoS requirements. Finally, we designed a management (reservation and scheduling) scheme for substreams to support better perceptual video quality such as the bound of end-to-end jitter. We have shown that our JSCC scheme is better than two other two popular techniques by simulation and real video experiments on the TCP/IP environment.

A well-suited approach to calculate the fractal dimension of digital images stems from the power spectrum of a fractional Brownian motion: the ratio between powers at different scales is related to the persistence parameter H and, thus, to the fractal dimension D=3-H. The signal-dependent nature of the speckle noise, however, prevents a correct estimation of fractal dimension from synthetic aperture radar (SAR) images. Here, we propose and assess a novel method to obtain D based on the multi-scale decomposition provided by the normalized Laplacian pyramid (LP), which is a bandpass representation obtained by dividing the layers of a LP by its expanded base band and is designed to force the noise to become signal independent. Extensive experiments on synthetic fractal textures, both noise free and noisy, corroborate the underlying assumptions and show the performances, in terms of both accuracy and confidence of estimation, of pyramid methods compared with the well-established method based on the wavelet transform. Preliminary results on true SAR images from ERS-1 look promising as well.

A system has been designed with the aim of helping the communication between a producer of holographic stereograms and users of three-dimensional (3D) computer graphics. The user of the system communicates with the producer through a small program, an applet, which is transferred over the Internet. In the applet the user makes all settings necessary for producing the hologram, and then sends the 3D file together with its settings to a holographic printer for production of the hologram. The system uses virtual reality modeling language as an interchangeable graphics format and Java as programming language. It is believed that the system will significantly improve the dissemination of holographic hard copies to ordinary users of computer graphics.

The image reference approach is very popular in industrial inspection due to its generality for different inspection tasks. Unfortunately, this approach is sensitive to illumination variations. A novel illumination compensation algorithm is proposed in this paper for correcting smooth intensity variations due to illumination changes. By using the proposed algorithm as a preprocessing step in the image reference based inspection or localization, we can make the image inspection or localization algorithm robust against spatially smooth illumination changes. This technique is very useful to achieve a reliable automated visual inspection system under different illumination conditions. The proposed illumination compensation algorithm is based on the assumption that the underlining image reflectance function is approximately piecewise constant and the image irradiance function is spatially smooth. Reliable gradient constraints on the smooth irradiance function are computed and selected from the image brightness function by using a local uniformity test. Two surface fitting algorithms are presented to recover the smooth image irradiance function from the selected reliable gradient constraints. One is a polynomial surface fitting algorithm and the other is a spline surface fitting algorithm. The spline surface fitting formulation leads to solving a large linear system, which is accomplished by an efficient preconditioned conjugate gradient algorithm. Once the image irradiance function is estimated, the spatial intensity inhomogeneities can be easily compensated. Some experimental results are shown to demonstrate the usefulness of the proposed algorithm.