Search

Digital scene matching algorithms have been used in both military and commercial image processing systems for years. The trend toward using multiple sensors in military imaging systems has generated anew interest in real time techniques to accomplish sensor fusion tasks such as field of view alignment. This thesis analyzes methods presently in use and introduces a novel algorithm that improves scene correlation performance. The focus of the new technique is in the segmentation area, where significant features are extracted from background and clutter. These performance improvements are especially helpful when the scene contains excessive noise and or lacks detail, a trouble spot for standard correlation systems. The restrictions imposed on the system design include implementations possible for real time processing and a minimum of hardware and power consumption. Simulations of the algorithms programmed for an image processing board by an IBM personal computer are discussed.

A visual flight simulator requires a huge amount of image data to be stored in the database. To make a photo-based system feasible an image compression scheme must be devised to compress the data. An adaptive discrete cosine transform (DCT) technique is used to compress 24 bit color images to an average of 3 bits per pixel. The bits for the image are distributed based on the relative activity in different parts of the scene. Results and error analysis are presented.

A simulation of an alternate implementation of a redundant busing network based on the Teradata Ynet architecture is presented. An overview of the Teradata DBC/1012 data base parallel processing computer including the Ynet, an active logic busing network, is given. Other multiprocessor busing networks are examined and compared to the standard Ynet and the alternate Ynet. In the standard Ynet system, two networks, called Ynets, process message packets concurrently. When one of the Ynet paths fails, the system is reset. The remaining Ynet path restarts using the previously interrupted packets and processing continues without the aid of the failed Ynet. In the implementation presented here, the two busing networks process the message packets in parallel. Now, when one of the Ynet paths fails, the other continues processing the packets without interruption. This implementation can be referred to as a parallel Ynet. The advantages and disadvantages of the parallel Ynet are discussed and suggestions for further research are given. Listings and sample outputs are included in the appendices.

The Hough transform silhouette identification method requires a consistency of edge direction in the identification of similar silhouettes. Many gradient operators used in Hough preprocessing require a thresholding and a non-maxima suppression routine to aid the localization process. These routines may delete edges or cause edge fragmentation. These anomalies degrade the Hough performance due to the lack of accurate silhouette extraction, and reduce the correct localization in the Hough accumulator. Noise or sampling errors can be removed by several enhancement routines presented. They are mean, median, symmetric nearest neighbor, hi pass, and low pass filters. An edge detection process is presented which produces a directional image and the confidence image that allows subsequent image analysis the ability to determine if the detected original edge orientation is accurate, and to what degree. The orientation confidence is produced by comparing a 7 by 7 operator and the Compass Gradient operator. This allows the Hough process the ability to modify the position of accumulation, thereby improving the Hough localization process.

Cartography, the art of producing maps, is an extremely tedious job which is prone to human error and requires many hours for the completion of maps and their digital data bases. Cartography is a classic example of a job that needs to be automated. Through the new advances in image processing and pattern recognition, the automation of this task is made possible with the cartographer acting as a supervisor. This paper reviews current cartographic techniques, and examines design considerations for a fully automated cartographic system. The benefits of such a system would be improvements in speed, flexibility, and accuracy. The role of the cartographer, with such a system, would change to a process supervisor rather than that of a mass data entry.

Psychological testing has shown that there is an early preattentive stage in the human visual system. At this level, simple features and properties of objects known as visual primitives are detected spatially in parallel by groupings of cells in the visual cortex known as feature maps. In order to study this preattentive stage in a machine vision system, the biologically inspired, highly parallel architecture of the artificial neural network shows great promise. This paper describes how the unique architecture of the counterpropagation neural network was used to simulate the feature maps which detect visual primitives in the human visual system. The results of the research showed that artificial neural networks are able to reproduce the function of the feature maps with accuracy. The counterpropagation network was able to reproduce the feature maps as theorized, however, future research might investigate the abilities of other neural network algorithms in this area. Development of a method for combining the results of feature maps in a simulation of full scale early vision is also a topic for future research that would benefit from the results reported here.

Striving to achieve the cognitive power of the human brain, researchers have turned to studying Artificial Neural Network (ANN) models through computer simulation. Computer programs have been written that permit the testing and evaluation of different ANN models. These programs provide facilities to layout different networks using a limited set of models, but do not allow the user to easily develop and create new training algorithms or activation functions. ANNview was written to provide the neural researcher with a Computer Aided Software Engineering (CASE) environment to develop new ANN models and a Graphical User Interface (GUI) to create and interconnect processing elements. The provided tools permit users to easily generate new ANN and processing element models from existing models. These models can then be tested and evaluated in different networks layouts using the existing simulation program simANNs and its parallel extension, parsimANNs. This study shows the benefits and power of using a CASE tool in developing and evaluating ANNs.

In the past decade, the use of control and diagnostic reasoning systems in different areas of government, industry, and university operations has increased. A great number of these systems find their basis in engineering, specifically in process control. The majority of the time devoted to the development of these systems is spent in the areas of Knowledge Engineering (KE) and Knowledge Acquisition (KA). Extensive research for the development of systems that perform the KE task is underway. The work presented here is an approach toward automatic knowledge acquisition. The objective of this research was to construct a complete knowledge base for a diagnostic and control reasoning system from information that resides in Computer Aided Design (CAD) databases. The purpose of CAD is to electrically order the components of a physical system under the direction of a human operator. However, the majority of designs in CAD databases do not contain any knowledge about the flow and operation of the systems they represent; they are simply electronically-stored drawings. It is also possible that the information that does reside in the CAD system is not complete or accurate enough to directly generate a viable knowledge base for an automated reasoning system. An algorithm was developed that automatically generates missing information and corrects potential mistakes made by the CAD draftsperson by using novel constraint labeling techniques. This work will decrease the amount of time spent in the generation of the knowledge base for diagnostic reasoning systems. It will also make possible the creation of more reliable knowledge bases since less hand coding will be required. Because validation and verification are inherent in the automated knowledge acquisition process, it is possible that this work could be expanded to be used as a designer's aide to monitor the draftsperson's activities during CAD drawing development.

Spatial transforms play a central role in image and signal processing applications. The Walsh-Hadamard Transform, due to its simple integer basis functions, provides superior computational simplicity and minimal memory requirements. In retrospect, the Fourier Transform, while providing a powerful tool for image and signal processing, has a computational complexity that makes it undesirable for applications where computational time and memory space are limited, such as with most fine-grained architectures. An objective of this research was to develop a fast parallel algorithm of the two-dimensional Walsh-Hadamard Transform for a fine-grained parallel machine - The Geometric Arithmetic Parallel Processor (GAPP) was chosen as the implementation platform. Different fast parallel algorithm techniques are investigated and the optimal two-dimensional Fast Walsh-Hadamard Transform (FWHT) is derived by combining these techniques. A comparison between the Walsh-Hadamard Transform and the Fourier Transform is given and special emphasis is placed on the advantages and disadvantages of these two transforms when performed in a fine-grained parallel architecture. Possible applications of the Walsh-Hadamard Transform in the signal and image processing fields are discussed.

One approach to model based computer vision as used for recognition is to store a database of wireframe models and then compare these to some digitized image of a scene. For even a single object, how one should effectively match an object model to an image remains an open question. We restrictively define recognition as determining the presence or absence of an object, and determining its pose-position and orientation fixed by six spatial parameters of translation and rotation. These parameters are incorporated into a camera model which sets the viewpoint of an observer and obtains a 2-D perspective projection of the object model. One objective is to efficiently position the camera such that the model's projection coincides with an object's digitized image- an object of unknown pose. Several researchers have described the above as an optimization problem and have proposed various deterministic solutions which are all somewhat limited in effectiveness. We propose a stochastic method which considers camera positioning as a Markov process, hence allowing simulated annealing to be applied for optimization. The objective function is a least squares distance between model projection and input image vertices. In addition, to speed up the search we employ the characteristic view concept. A "qualitative" match is first performed by checking that the number of visible object faces agree between model and image. An underlying philosophy (inspired by the Marr paradigm) that long term computer vision advances will come mainly from biologically plausible algorithms spawns discussions that examine links between human and machine vision.

The topic of supervised learning within the conceptual framework of artificial neural network (ANN) models is addressed. An ANN is a parallel distributed processing system that consists of many computationally simple processing elements interconnected through unidirectional weighted connections. Such networks, which are roughly patterned after biological nervous systems, have been proposed for use in areas in which the traditional von Neumann computer architecture has been relatively unsuccessful. Learning in these networks is accomplished through the use of algorithms that adjust the values of the connection weights. The work presented here addresses the issue of improving the rate at which ANNs can learn to achieve the mapping of an input pattern to a desired output pattern. The most successful learning algorithms for accomplishing this task are based on gradient descent error minimization techniques. However, the large amount of training time that such algorithms require is currently one of the factors that limits the use of ANN systems. A number of improvements to existing algorithms were developed and analyzed through the use of simulation. Furthermore, a highly flexible simulation environment based on an object-oriented software paradigm was developed and successfully used to implement and test these novel learning algorithms. A comparison of this object-oriented approach with the traditional techniques used by existing ANN simulation systems is performed. Finally, the issue of simulating ANN models on parallel hardware is considered. An associative memory ANN was implemented on a systolic array architecture. The memory requirements and processing times of this implementation are analyzed based upon the number of processing elements required by the ANN and the number of training patterns. Compared with other digital implementations, this design is shown to yield significant improvements in runtime performance and offers the capability of using ANN associative memories in real-time applications.

A need was recognized for the development of a hardware independent simulator for the Geometric Arithmetic Parallel Processor, or GAPP. This simulator would allow the user to test algorithms targeted to GAPP systems using readily available personnel computers. The development of the simulation, a GAPP Instruction Simulator, or GIS, is the basis of this research report. This report includes a discussion of earlier parallel processors of similar type to the GAPP, a discussion of the GAPP's operation and instructions, and illustrations of GAPP applications. The report focuses on the GAPP Instruction Simulator with a discussion of its operation and a listing of the software implementation. The GIS software was compiled and test executed with results discussed in the conclusion of this report. the GIS provides a useful simulator for GAPP application design testing without having GAPP hardware available.

A compact neural network architecture using a hybrid digital-analog design is implemented in Very Large Scale Integration (VLSI). Various ways to implement an efficient and practical neural node in electronic hardware are examined. Digital, analog and a combination of these methods are evaluated as a means of implemention. Design examples of these methods using VLSI are presented and the advantages and disadvantages are discussed. The network architecture presented uses VLSI to alleviate some of the interconnectivity problems involved with current neural network implementations. The resulting chip can be used to perform simulations in real time of different learning algorithms and to build large systems of neural nodes. The different node models were designed and tested using VLSI computer Aided Design (CAD) tools. A complete network was then constructed in VLSI using the most efficient and practical node model.