A review of computer architectures for machine vision applications is presented. Pipeline and cellular architectures for raster-to-raster and raster-to-list operations are described. Advantages and disadvantages of the various architectures are discussed.

The technology of machine vision systems has matured rapidly over the past four years. Over 100 small companies have brought products to market aimed at putting "intelligent eyes" on the factory floor. In the early days of the machine vision industry, many different technological approaches were tried. The first practical systems were based on binary images which were processed by a frame grab, window, and pixel count. The SRI algorithms extended the binary technology to the point that object recognition and location was a possibility. It quickly became clear that this type of technology was not robust enough for dependable operation on the factory floor. Customers demanded systems that could operate for months at a time unattended, even if the intensity of light changed or the fixturing of the objects became sloppy. In addition, potential customers were attracted to the technology of machine vision, only to find out that processing speeds were too slow.

This paper is an in-depth look at determining the location and quality of image features by normalized correlation. We have investigated its accuracy, repeatability, and susceptibility to various forms of image degradation, and compared the results to other methods of locating features. We also present a novel, inexpensive device that performs the necessary computations very efficiently. The emphasis is on locating features for current, practical applications in manufacturing automation.

A large number of computer system designs for image analysis have been proposed, and many have been or are being constructed. Although many of them share similar features, the use of proprietary terminology and lack of detailed standardized processing examples in descriptive documentation makes understanding and comparison difficult. Consequently, the a priori estimation of algorithm performance on a given system becomes a combination of ad hoc guesses and somewhat idealized extrapolation. Also, in concentrating on only a portion of the overall data flow, many of these designs have optimized only a part of the system and have reduced the overall price/performance ratio. This paper identifies the fundamentally distinct primitive types of image transformations used in image processing and discusses the characteristics of typical applications and missions. Next, a taxonomy of architectures is presented which enables the various approaches to be unambiguously categorized and evaluated. Several existing designs are used as examples, and the relative merits of each are discussed.

The paper gives an hyperbrief review of computer vision, concen-trating on representative achievements in early vision. The account attempts to stress the underlying unity of its scientific foundations and intellectual achievements.

Computer recognition and inspection of objects is, in general , a complex procedure requiring a variety of kinds of steps which successively transform the iconic data to recognition information. We hypothesize that the difficulty of today's computer vision and recognition technology to be able to handle unconstrained environments is due to the fact that the existing algorithms are specialized and do not develop one or more of the necessary steps to a high enough degree. Our thesis is that there are no shortcuts. A recognition methodology must pay substantial attention to each of the following five steps: conditioning, labeling, grouping, extracting, and matching.

Many types of image processing operations can be performed sequentially at frame rates, but many of the global operations needed in computer vision systems cannot be performed in real time unless suitable parallel hardware is available. This paper considers an important class of such operations, involving image segmentation by detection and extraction of global regions and features, and describes algorithms for carrying out these operations on parallel "pyramid" hardware.

Hypercube architectures are introduced. The reasons behind their becoming the first widespread commercial massively parallel processors are outlined. A classification for image pattern recognition is proposed and characteristic algorithms are presented. Some simple ideas for organizing these algorithms to execute efficiently on hypercubes are also presented. The current programming model for hypercube machines is explained and illustrated with one of the characteristic algorithms. Preliminary performance figures are discussed for the NCUBE/six, a 84 processor commercial hypercube, followed by some concluding remarks.

Optical Pattern Recognition has provided many attractive algorithms and architecture for advanced use in Automatic Target Recognition (ATR) and computer vision. This work is reviewed and highlighted in this paper. Attractive aspects of all of this research are: its attention to distortion-invariant, multi-target object recognition and the extensive testing which has been performed of these various architectures on large databases, as well as the design and fabrication of several quite compact optical processing architectures. Recent Artificial Intelligence (AI) techniques promise to further advance optical processing. These issues are summarized herein.

The success of a variety of techniques in pattern recognition led to a sense that pattern recognition was "a solved problem" and a mature discipline. Simultaneously, the appearance of methods of Artificial Intelligence (AI) seemed to offer hope for the solution of many of the problems that pattern recognition techniques had not adequately dealt with. In particular, there has been interest in the development of various capabilities of expert systems and knowledge-based-systems for extending the scope of pattern recognition. One application where this has been very prominent is the analysis of images, i.e., computer vision.

A critique is provided which first analyzes the nature of image processing (machine vision) algorithm development. The critical need for, and the pre-requisite techniques for, addressing improved productivity of algorithm development are discussed. Results and directions of development toward a high-productivity "vision solution factory" are discussed.

Researchers with backgrounds in fields other than pattern recognition are becoming increasingly aware of the contribution that quantitative image analysis and pattern recognition can make in their work. These unlimited application possibilities hold a bright future for the technology of pattern recognition. Required, however, is an easy to use, simple to understand user interface to the system. This paper discusses those requirements and presents a classifier design example that was developed on one such system. Also included are a glossary and a summary of the mathematics behind the Bayes classifier.

After about twenty-five years of growth, image analysis technology has advanced to the point where the design of machines that model real scenes in real time is at the technological frontier. Consequently, we see the beginning of a merging of the technologies of image analysis and automatic control. This merger is leading to the development of complex distributed systems in which decisions and actions are based on realistic models of possibly time-varying scenes. We refer to this merged technology as vision automation. In the past decade, reflecting the impact of artificial intelligence, image analysis has expanded to scene understanding, and automatic control has evolved into intelligent control. Thus, vision automation consists of two major components: scene understanding and intelligent control. (We prefer the term "scene understanding" to the more commonly used "image understanding", since we want to understand the scene, not the image.)

A description is given of a prototype address block finding system which is currently under development. This system is intended to support a wider application of machine vision and robotic technologies for the handling of mail. Monochrome digital images of mail are subjected to a sequence of binarization, clustering, and ranking algorithms which are designed to automatically determine the position, extent, and orientation of the destination address block. The prototype system is implemented on a LISP machine to facilitate the extension of the complexity of feature interpretation on an Expert System. Success rates for the end-to-end process were tested on 40 mail images representing a wide variety of mail characteristics.

This paper concerns a critique of several line-drawing pattern recognition methods such as picture descriptive language[15], Berthod and Maroy's methods[1], extended Freeman's chain code [4,22], tree grammar[5-6] and array grammar[3] [20,21]. Then a new character recognition scheme using improved extended octal code as primitive is introduced, which has certain advantages such as flexible size, orientation, variations, fewer learning samples needed and lower degree of ambiguity. Finally the concept of semantic pattern recognition is discussed.