This paper describes a method of optimal 3-D object surface representation used for feature- based parametric matching. First, the spatial and frequency domain methods are presented for the determination of an optimal grid spacing for 3-D object surface representation. The spatial domain approach is based on the measurement of the total variations of the object surface function, f(x,y), in the constant x and constant y planes. The frequency domain approach utilizes the Nyquist criterion and an ideal 2-D rectangular low-pass filter in an iterative operation. Two error criteria are defined based on the magnitude and shape features of f(x,y) as the measure of effectiveness of the optimal grid spacing for 3-D object surface representation. Second, a parametric matching is devised as a second order mapping function which maps the boundary features of f(x,y) in the cross section planes of the object and model surfaces.

Computer graphics rendering and the compression of range data with its associated color data impact the quality and speed of reconstructing photo realistic images from range/color scans of real world scenes. A method of consolidating orthographically scanned, quantized range data into polygons was developed with the intent of being both practical and efficient in its processing and storage of only the information required for producing a new photo-realistic image from nearly the same perspective as the original scan. Tests were performed on both actual range data and data extracted from the z-buffer of polygonal computer graphics renderings. In each case, for every range value a 24-bit color was associated with each range value. A polygon was constructed to represent the range and color data. Surface normals were calculated for each polygon, and consolidation of neighboring polygons was performed if their normals occupied the same quantized region of the Gaussian sphere. Results on polygonal consolidation ratios and reconstructed image quality as a function of quantization will be presented.

Two processes to produce simulated imagery are contrasted. The first is a physical approach based on terrain board model scenes, illumination sources, and an electro-optical camera. The second method utilizes computer graphics to render a mathematical model of a scene. Examples of the same scene modeled physically and by computer graphics are compared with respect to their utility in portraying image quality characteristics.

Our intent is to obtain images which most clearly differentiate soft tissue types in Magnetic Resonance Image data. We model the three unknown intrinsic parameter images and the data images as Markov random fields and compare maximum likelihood restorations with two maximum a posteriori (MAP) restorations. The mathematical model of the imaging process is strongly nonlinear in the region of interest, but does not appear to introduce local minima in the resulting constrained multidimensional optimization procedure. The application of non- quadratic prior probabilities however does require global optimization. We have developed a unique approach towards image restoration that produces images with significant improvements when compared to the original data. We have extended previous results that attempt to determine the intrinsic parameters from the MRI data, and have used these intrinsic parameter images to synthesize MR images. MR images with different TE and TR parameters do not require additional use of an MR scanner, since excellent synthetic MR images are obtained using the restored proton density and nuclear relaxation time images.

Fractal image compression is achieved by Fractal transformation of images. To complete a fractal transformation, one has to solve a Fractal equation. In this paper, we will develop an algorithm for solving Fractal equations of black and white images. We will first briefly derive several types of Fractal equations. Then we will present an algorithm for solving Fractal equations. We will discuss images with more than two pixel intensities in the second part of this paper.

Methods of multiresolutional image representation are well known and widely used in computer vision. They have many serious advantages because they provide compact encoding of images and quick search of the objects within the images. However, the traditional methods of multiresolutional image representation presume forming a set of copies of the image at different resolution levels with a constant resolution at each level and constant ratio between two consecutive levels which does not depend on the structure of the processed image. On the other hand, it has already been demonstrated that every image has a limited and innate spatial scale range as well as a limited resolution range, and the result of transformation depends on their initial selection. The main idea of the developed approach is not to form constant resolution levels but rather to grOw a resolution tree which depends on the structure of the image and the task at hand. The resolution tree grows sequentially during the topdown processing of the image. As a primary feature, we have selected an oriented edge segment (a stroke). The strokes are extracted with a particular resolution of the level at the points of local maxima of the brightness gradient. Strokes can have different thickness, length, and orientation. Each of them subsequently is decomposed into another set of strokes at a higher resolution, and so on. Thus, the whole gray-level image is being transformed into a hierarchical set of strokes characterizing forms of objects with different degree of generalization depending on the size of objects in the images. This method will transform the original image to a hierarchical graph which allows for efficient coding in order to store, retrieve, and recognize the image. The method which is proposed is based upon finding the resolution levels corresponding to each image, and each subset of the image individually which minimizes the computations required. This becomes possible because of the use of a special image representation technique called Multiresolutional Attentional Representation for Recognition. This feature turns out to be efficient in the process of finding the appropriate system of resolutions and construction of the relational graph. The process is performed by a three-layer neural network (NN) with intra-layer interactions between neurons, the receptive fields of which are selectively tuned to detect the orientations of local contrasts in parts of the image with appropriate degree of generalization. The NN forms a stroke-sketch of corresponding resolution for each partition of the image.The scale parameter characterizing the resolution of processing is assigned to NN from outside. In order to correct the stroke-sketch formed by the NN, the bottoin.up algorithm is used to complete the stroke-sketch with missing strokes. Three supportive algorithms make possible this top-down/bottom-up operation. The Resolution Level Estimation (RLE) Algorithm for apriori evaluation of the preferable scale parameter for the whole image and all of its partitions. The Hierarchical Image Decomposition (HID) Algorithm finds partitions. Finally, the Algorithm of Corrections by Gestalt Heuristics (CGH) provides the "gestalt" unity of the strokes that may have been "damaged" during processing.

Several theorems regarding to Gaussianity, independency, correlation, stationary, ergodicity, autocorrelation function and spectral density function, etc. of the computed image are presented in this paper. These theorems are focused on one issue-the spatial correlation among the pixels. The applications of these theorems to image analysis are also discussed.

We present several objective functions for vector field segmentation. Leclerc's MRF model is extended by the addition of information-theoretic penalties for regions and distinct means. Massively parallel optimization algorithms are presented. Standard methods of signal detection and estimation are used to develop a theoretical performance analysis, which quantitatively predicts the performance at realistic noise levels. Theoretical and experimental results agree fairly well.

The use of mathematical morphology in low- and mid-level image processing and computer vision applications has allowed the development of a class of techniques for analyzing shape information in color images. These techniques have shown to be useful in image enhancement, segmentation, and analysis. In this paper, we develop and test scalable parallel algorithms necessary to implement a class of morphological filters on a parallel computer, specifically, the MasPar MP-1. We examine the issues relative to the parallel implementation of the algorithms and show that real-time enhancement of high resolution color images is possible.

We present a technique for Image Segmentation using Neural Tree Networks (NTN). We also modify the NTN architecture to let is solve multi-class classification problems with only binary fan-out. We have used a realistic case study of segmenting the pole, coil and painted coil regions of light bulb filaments (LBF). The input to the network is a set of maximum, minimum and average of intensities in radial slices of a circular window around a pixel, taken from a front-lit and a back-lit image of an LBF. Training is done with a composite image drawn from images of many LBFs. Each node of the NTN is a multi-layer perceptron and has one output for each segment class. These outputs are treated as probabilities to compute a confidence value for the segmentation of that pixel. Segmentation results with high confidence values are deemed to be correct and not processed further, while those with moderate and low confidence values are deemed to be outliers by this node and passed down the tree to children nodes. These tend to be pixels in boundary of different regions. The results are favorably compared with a traditional segmentation technique applied to the LBF test case.

In this paper a new algorithm to estimate dense displacement fields from a sequence of images is developed. The algorithm is based on modeling the displacement fields as Markov Random fields. The Markov Random fields-Gibbs equivalence is then used to convert the problem into one of finding an appropriate energy function that describes the motion and any constraints imposed on it. Mean field annealing, a technique which finds global minima in nonconvex optimization problems, is used to minimize the energy function, and solve for the optimum displacement fields. The algorithm results in accurate estimates even for scenes with noise or discontinuities.

A system has been developed which performs adaptive color image enhancement for use in a graphic arts prepress environment. This system uses a model of human vision and provides independent control of sharpness, grain noise, light and dark halos, brightness, contrast and dynamic range. The sharpness, noise reduction and halo control processes are based on adaptive subband filtering techniques which exploit the stationary characteristics of film grain and the non-stationary nature of image information. The contrast and dynamic range processes mimic the human visual perception of illuminance and reflectance in image subject matter.

Itek Optical Systems has developed a hybrid multispectral imagery simulation capability based on physical images of a terrain board and computer modeling of radiation propagation. This process produces multispectral imagery within the 0.4 micrometers to 2.5 micrometers wavelength region that reflects the complex interactions among the ground scene reflectance, atmospheric radiance and attenuation, system acquisition conditions, and sensor performance characteristics. This process is used to evaluate performance of multispectral sensor designs by comparing output products representative of each design. Imagery produced by this process is also well suited for automatic processing algorithms because various imaging parameters are easily and independently altered in a controlled manner. These parameters may include spectral band placement, bandwidth, number of bands, image spatial resolution, sensor noise, feature location, and mixed pixel composition. The resulting simulation imagery can be otherwise identical in scene content, allowing direct analysis of algorithm performance as a function of specific input scene and acquisition conditions.

In this paper, a neuromorphic multilayer architecture, called KYDON, is presented. In particular, the structural design of the layer nodes and the low and high level vision tasks performed by KYDON nets are described. KYDON architecture has 'k' layers of nodes connected in full hexagonal mesh connectivity. The lowest layer capture images from the environment by employing 2-D photoarray. The top most layer deals with image interpretation and understanding. The intermediate layers perform various process to bridge the bottom most layer to the top most layer. KYDON use graph to represent the knowledge, extracted from the image. An important feature of KYDON is that KYDON does not have any host computer or control processor to handle I/O and other miscellaneous tasks.

Processing gray-scale realizations of images that are ideally binary (such as gray-scale realizations of printed characters) is problematic due to the fact that gray-scale processing should be consistent with the binary nature of the ideal image. Essentially, any final decision (such as the recognition of a specific character at a specific location) should reflect the content of the ideal image, which is generally unknown. Too often, a gray-scale realization of an ideal binary image is processed using methods appropriate for gray-scale realizations of ideal gray- scale images. These should not be expected to lead to decision procedures appropriate for binary images. Fuzzy morphological algorithms do not assume probabilistic knowledge of the degradation process; however, they mirror the processing that one would have performed were the ideal binary image known. Thus, they lead to decision procedures consistent with those that would have been taken following processing of the ideal binary image. In this paper we discuss the fuzzy hit-or-miss transform based shape detectors that are capable of detecting geometric shapes in the presence of considerable additive as well as subtractive random noise. There exists an infinite number of realizations of this shape detector and the determination of which detector is suitable is application dependant; nevertheless, there exist a general set of heuristics for selecting the appropriate realization. We also carry out extensive noise- sensitivity analysis for a few of these shape detectors.

This paper studies an object recognition problem, that is, the problem of determining whether a given perspective image is obtained from a 3-D object to be recognized or not. As an extension of Ullman and Basri's approach, it is found that any perspective image of an object can be expressed as a certain type of nonlinear combination of four appropriate perspective images of the same object. We show that any image of an object with not only a rigid 3-D transformation but also a nonrigid transformation has this property. In order to recognize a 3- D object, we have only to store four perspective images and, whenever a new perspective image is given, determine whether it can be expressed as a combination of the four images. This implies that we no longer need to recover the 3-D information of an object explicitly under perspective projection. Our investigation shows that four perspective images have sufficient information to recognize a 3-D object.

In an Image Understanding framework, our aim is to reconstruct an actual indoor scene from a (sequence of) color pair(s) of stereoscopic images. The desired (synthesis-oriented) description requires the analysis of both 3D geometric and photometric parameters in order to use the feedback provided by image synthesis to control the image analysis. The environment model is a hierarchy of polyhedral 3D objects (planar lambertian facets). Two main physical phenomena determine the image intensities: surface reflectance properties and light sources. From illumination models established in Computer Graphics, we derive the appropriate irradiance equations. Rather than use a point source located at infinity, we choose instead isotropic point sources with decreasing energy. This allows us to discriminate small irradiance gradients inside regions. For indoor scenes, such photometric models are more realistic, due to the presence of ceiling lights, desk lamps, and so on. Both a photometric reconstruction algorithm and a technique for localizing the 'dominant' light source are presented along with lighting simulations. For comparison purposes, corresponding artificial images are shown. Using this work, we wish to highlight the fruitful cooperation between the Vision and Graphics domains in order to perform a more accurate scene reconstruction, both photometrically and geometrically. The emphasis is on the illumination characterization which influences the scene interpretation.

The problem of realistic, high-resolution, earth surface representation for real-time, rendered, video-quality perspective view generation has been approached by using a quantized rendering transform to code image measurements into physical surface modeling descriptors. This paper describes a physical earth surface model and approximates natural light energy scattering equations to derive a transform between photogrammetric measurements and model parameters. This transform was used to translate a 12 Gbyte, photo image, data base covering 400 sq km of Ft. Hunter Liggett, CA into a 1.2 Gbyte surface descriptor file. A prototype transputer-based parallel processing system is also presented. The system uses the rendering transform to calculate real-time perspective views at operator-selectable times, seasons, and environmental conditions. The system produces video-realistic perspective views at a rendering rate of .5 Mpixels/second and is scalable by a factor of 80.

An iterative estimation procedure is applied to the determination of 3-D parameters from Remote Sensing image data. The 3-D parameters describe the geometry of the scene viewed. The scene consists of man-made objects, such as buildings. The estimation procedure is based on a prediction-correction mechanism. The prediction is generated using a model of the imaging process. An essential point in such an estimation algorithm is the comparison level used. The choice of the comparison level is dependent on many factors. The key factors studied are the possible information loss in data reduction, and the introduction of modelling errors in the prediction model. These modelling errors might be smaller on higher data levels. Two comparison levels are explored in depth in this research. The first comparison level studied is the level that describes the shape of regions found by segmentation. The predicted and measured regions are matched using an error measure based upon the regions run-length encoding. The error is minimal if length and position of the scan line segments match. The second level studied is reached by calculating the moments of the regions. The error measure consists of the subtraction of predicted and measured moments. Results show that the proposed approach is possible. The speedup obtained, compared to using the raw data as comparison level, is substantial, typical in the same order as the data reduction applied. The accuracy of the final parameter values is comparable between the tested levels.

The development and evaluation of Multiple-View 3-D object recognition systems is based on a large set of model images. Due to the various advantages of using CAD, it is becoming more and more practical to use existing CAD data in computer vision systems. Current PC- level CAD systems are capable of providing physical image modelling and rendering involving positional variations in cameras, light sources etc. We have formulated a modular scheme for automatic generation of various aspects (views) of the objects in a model based 3-D object recognition system. These views are generated at desired orientations on the unit Gaussian sphere. With a suitable network file sharing system (NFS), the images can directly be stored on a database located on a file server. This paper presents the image modelling solutions using CAD in relation to multiple-view approach. Our modular scheme for data conversion and automatic image database storage for such a system is discussed. We have used this approach in 3-D polyhedron recognition. An overview of the results, advantages and limitations of using CAD data and conclusions using such as scheme are also presented.

This paper makes a short description of the visualization bridging the gap between man and the rich computer environment.