processing

This project investigates experimental methods to produce a particulate reinforced material consisting of solid particles embedded in an ice matrix. The objective is to develop a new material that can be used to conduct ultrasonic nondestructive testing of complex shape engineering components such as those produced with additive manufacturing methods

The research being conducted under this contract has two main themes: a) Parallel algorithms for image processing b) Expert systems for aerial image understanding A supplemental theme is the mathematical modelling of images, with emphasis on texture modelling.Thirteen technical reports were issued on the contract; abstracts of these reports are appended to this Annual Report.Numbers in brackets in the following paragraphs refer to this list of abstracts.
It has been well known for over 25 years that meshconnected cellular arrays are a very natural architecture for image processing at the pixel level.We have studied methods of increasing the reliability of cellular arrays by sharing data and computations between adjacent processors in the array [l].
More general classes of SIMD (single instruction stream, multiple data stream) architectures have also been studied on this project.We have studied methods of allocating data among processors in an SIMD system so as to perform parallel inferences using minimum computation and data transfer time [2].
A comparative survey of software systems for image processing is given in Several of our studies have concerned the mapping of networks of processors onto other networks.In particular, we have shown how to embed trees and arrays into hypercubes efficiently (i.e., without great increase in number of nodes or in distance between neighboring nodes) [10].We have also partially supported the research of Prof. I. V. Ramakrishnan [5,6,11], who is studying the mapping of various types of algorithms onto linear arrays of processing elements which are well suited for VLSI implementation.

Cbef, TecnbAal Information Divigion
In the area of expert image understanding systems, we have developed a control structure that deals with objects satisfying given spatial relations.By accumulating predictions originated from existing instances, our system can dynamically reason about what to do in order to construct interpretations of the image.The system is being applied to a relatively simple domain consisting of images of houses and roads on aerial photographs [4,13].
In connection with image modelling, some excellent work on the computer synthesis of natural textures was done by Drs.Andre Gagalowicz and Songde Ma on a six-month visit to our laboratory from INRIA (France).They developed both parallel [7] and sequential [8] algorithms for synthesizing textures having given second-order spatial statistics.An important new result [9] was the development of methods for directly synthesizing textures on curved surfaces, rather than first synthesizing them on a plane region and then mapping that region onto the surface.This work has important applications to the fast synthesis of textures for the dynamic simulation of natural scenes.On a more abstract level of image modelling, we have investigated methods of defining fuzzy relationships among image parts, with emphasis on relationships of adjacency and surroundedness [121.

MANN
Appendix: Abstracts of Technical Reports
ABSTRACT: This paper describes several methods for increasing cellular array reliability for image processing through redundancy, with emphasis on error masking methods tnat preserve continuous operation of the array.We define n-module data sharing * (NMDS) for a cellular array, which achieves redundancy by sharing data and computations between adjacent processors in the array.
The time, space and reliability performance of NMDS is compared to standard n-module redundancy (NMR).
ABSTRACT: Let S=Jl'S21* ... IS be a set of data objects, and f be function from SxS into q.For Tc{l,2,.. .,N,}2 the operation Tf applies f to every element of { (Si,Sj)1I(i,j) ET~c SxS.For the optimal computation of Tf on a SIMD system with p processors, the data allocation problem consists in distributing objects to the p processors such that the computation and data transfer time is minimized.In the paper, the cases T=tl,2,...,N} 2 and T={(i,j)1l!Ej<i--N} are dealt with.

3.
Klaus ABSTRACT: We describe a control structure for building an Iage Understanding System.This system can deal with objects wth diverse appearances when consistent spatial relations exist beteenobjects.By accumulating consistent predictions origiaedfo existing instances, our system can dynamically reasnabout what to do in order to construct interpretations of the image.In this paper, we have discussed parts of the proposed system -the representation of spatial knowledge, the accumulation of evidence, the focus of attention mechanism, and the integration of constraints for top-down control.
ABSTRACT: This paper presents a formal model of linear array processors suitable for VLSI implementation as well as graph representation of programs suitable for execution on such a model.A distinction is made between correct mapping and correct execution of such graphs on this model and the structure of correctly mappable graphs is examined.The formalism developed is used to synthesize algorithms for this model.
ABSTRACT: A matrix-multiplication algorithm on a linear array using an optimal number of processing elements is proposed.The local storage required by the processing elements and the I/O bandwidth required to drive the array are both constants that are independent of the sizes of the matrices being multiplied.The algorithm is therefore modular, that is, arbitrarily large matrices can be multiplied on a large array built by cascading small arrays.The array is well-suited for VLSI implementation.
ABSTRACT: This paper deals with an optimization technique applied to natural texture synthesis.We propose a definition of a global criterion which is the mean square error between the statistical features of a natural original texture and those of an artificially generated one.A gradient algorithm is used to minimize this criterion.The statistical feature vector used was the autocorrelation function although this is by no means the only choice.The textures generated are very similar to the original ones.This method can be implemented in a highly parallel manner.
ABSTRACT: This paper presents a new method for the synthesis of textures on 3-D surfaces.To our knowledge, one basic technique has been presented up to now in the literature (see . In this standard method, textures are synthesized by mapping a rectangular template onto the curved surface.This method is complex, requires substantial computing time, and presents some drawbacks such as the possibility of obtaining aliasing effects and continuity problems along the edges of the curved templates.Procedures to eliminate these problems are available [11,12] but make this synthesis even more unattractive.The method proposed in this paper does not present te former drawbacks.We do not use a template mapping, which is a drawback in itself.The synthesis is achieved continuously on the surface, so that there are no edge effects and also no aliasing effects.This method is a simple extension of a procedure that we have proposed before in the literature [2,3] for planar textures.Any kind of texture can be reproduced with a good similarity to the reference texture used.It also has the important advantage that only one set of second order statistics (a small amount of data) needs to be computed on a planar version of the reference texture to synthesize this texture on any surface and at any distance.Some results on simple surfaces are displayed (cylinder, sphere), but the method holds for any surface and is relatively quick and easy.
ABSTRACT: A new method for the generation of natural textures is presented.Using a priori-given second order statistics (second order spatial averages or autocorrelation parameters) of a natural texture as input, we give a procedure to synthesize an artificial texture field in such a way that its second order statistics are equal to the deired ones.The synthesis is achieved directly without inventing higher order statistics, as was the case in earlier publications [3][4][5][6][7]15].This method allows us to synthesize gray tone texture fields while "controlling" their second order statistics in rather large neighborhoods.The synthesized textures are very similar visually to the original natural textures used to compute the second order statistics, but second order spatial averages give better results than autocorrelation parameters.This seems to strongly support the conjecture [6,8,101 that the visual system is only sensitive to the second order spatial averages of a given texture field, so that these statistics should be used to model textures.
ABSTRACT: The hypercube is a good host graph for the embedding of networks of processors because of its low degree and low diameter.Graphs such as trees and arrays can be embedded into a hypercube with small dilation and expansion costs, but there are classes of graphs which can be embedded into a hypercube only with large expansion cost or large dilation cost,