This paper presents design elements for a distributed image processing system. Requirements of large image processing systems suggest a configuration with tight clusters loosely coupled to other clusters. Within a cluster, user interfaces and resources are distributed according to security, redundancy, simplicity, expandability, cost, and timing requirements. All the various resources are controlled through layers of module similar to those of a network architecture.

In view of the 56 KBPS digital switched network services and the ISDN, low bit rate codecs for providing real time full motion color video are under various stages of development. Some companies have already brought the codecs into the market. They are being used by industry and some Federal Agencies for video teleconferencing. In general, these codecs have various features such as multiplexing audio and data, high resolution graphics, encryption, error detection and correction, self diagnostics, freezeframe, split video, text overlay etc. To transmit the original color video on a 56 KBPS network requires bit rate reduction of the order of 1400:1. Such a large scale bandwidth compression can be realized only by implementing a number of sophisticated,digital signal processing techniques. This paper provides an overview of such techniques and outlines the newer concepts that are being investigated. Before resorting to the data compression techniques, various preprocessing operations such as noise filtering, composite-component transformation and horizontal and vertical blanking interval removal are to be implemented. Invariably spatio-temporal subsampling is achieved by appropriate filtering. Transform and/or prediction coupled with motion estimation and strengthened by adaptive features are some of the tools in the arsenal of the data reduction methods. Other essential blocks in the system are quantizer, bit allocation, buffer, multiplexer, channel coding etc.

This paper describes an incremental transmission technique for displaying 3D imagery at different levels of detail and resolution. The objective of the technique developed is to allow a user to browse through imagery at a low resolution and zoom in only on specific features of interest. The technique described in this paper allows increasing the resolution and detail of the image by transmitting a relatively small amount of additional information and computing the new image locally in the display system. The need for systems with this capability is essential when (i) the resolution of the imagery exceeds the display resolution of the device, and (ii) a central file server on a network must transmit image-data to support several display devices.

A design of a freeze frame coder is presented. It is a considerable departure from normal full motion teleconferencing coders, as unlike the latter, freeze frame coders are less complicated due to the non-real time processing of the images. Typically, freeze frame coders find application in archival storage of images, freeze frame teleconferencing, etc.

A high quality color printing image compression method is presented. This contains dividing image into nxn blocks, vector-quantizing each block independently. In this stage, quantizing level is decided by means of range statistics. As the result of this metod, 1/11 of compression ratio and 30dB of S/N ratio is achieved.

A new method of destriping Landsat data and compensating for failed detectors is presented. The basic spatial frequency composition of the degradation is determined from an easily computed one-dimensional power spectrum of the image. A convolutional restoration kernel is then created from a modified version of the power spectrum.

An automated vision image analysis system to analyze images resulting from nucleic acid hybridizations is presented. Real-time image acquisition from autoradiographs is accomplished with a sensitivity to optical density to the degree that any spot visible by eye can be quantified by digital methods and enhanced by image processing techniques for observation. The system provides a tool for quantification of differences in gene expression, rapid processing of thousands of pairs of genes, storage in a database, search for small differences, large differences, and infrequently occuring genes. Background correction, normalization and pseudocolor "mapping" facilitate comparison and quantification directly from the CRT display. The image pairs, a percent change image and a difference image are presented in a quad display with coordinated spot density and change measurement simultaneously for each of the four images. The system provides a tool for analysis of altered gene expression in response to environmental stress, such as heat-shock and salt water in plant growth; hormone regulation, such as before and after hormone administration in animals; development; and carcinogenesis. The use of this system to study nucleic acid hybridization is contrasted with the study of 2-D protein electrophoresis gels for the determination of changes in gene expression and identification of corresponding proteins of significance.

The automatic determination of local similarity between two images (image registration) is one of the most fundamental problems of image processing and pattern recognition. A two-stage registration algorithm that is reasonably efficient and robust for rotational and translational displacement has been considered to determine relative shift between reference and search images. A new method of calculation for zero-order moment on a circular window is described, Based upon this measure, a computationally efficient first stage algorithm is obtained. On the second stage, Modified Circular Symmetric Autoregressive Random (MCSAR) model and invariant moments are used to determine a robust feature vector associated with the reference and subsearch images. Simulation results on real images are presented to evaluate rotation invariancy of the features and the performance of the proposed algorithm.

A three stage algorithm based on central moments and an adaptive classifier is developed to estimate the position of a noisy subimage with unknown angular misalignment within an image. The first stage derives the zero-order moment of pixels inside a circular window recursively and by comparing them to that of the reference image, a number of likely positions are selected. The second stage improves on the first stage by using several concentric circular windows. The third stage compares the higher order moments for the likely position obtained in the second stage. An adaptive classifier is used in parallel with the third stage in order to decide whether the computationally intensive high order moments should be calculated for a given neighborhood of the likely positions. This classifier ex-ploits the fact that the measure of similarity of the neighboring pixels does not vary substantially. Implementation of the algorithm on real images has indicated that, in many cases, the exact location of the rotated object within a noisy image can be determined.

A method for adaptively correcting for shading due to variations in illumination and imaging system imperfections, while at the same time reducing the number of gray levels in the original image on a pixel-by-pixel basis was developed. The two operations - shading correction and gray level reduction - are done in a single pass over the image and the local properties of the pixel being processed are utilized. Each scanline of the captured original image is processed independently enabling parallel processing of the scanlines to reduce the time to process the complete image. The technique described herein is applicable generally to any imaging system which is designed to work on reduced number of gray-levels. Through examples, it is shown that the resulting processed image has enhanced contrast, uniform background intensity, reduced number of gray levels (resulting in compression), and smoother edges (reducing staircasing) in multilevel images.

We have developed over the past few years new linear filtering methods for images which incorporate the use of properties of human vision in design. These methods apply to noise removal, image enhancement and image restoration. One interesting extension of our work is to the use of the anisotropy of vision to achieve better visual quality. In this paper we propose combining several of these processing tasks at once, by taking advantage of visual perception and of the image content information that can be estimated directly from the image. We present a brief overview of the image filtering techniques using properties of vision that we have developed and discuss the basis on which several processing tasks can be combined for the same image. We discuss briefly the approach taken to the implementation of such multistep algorithms on dedicated pipeline image processors.

In some flight simulation systems, the imagery is produced by continually applying perspective transformations on a few aerial photographs, which have been previously recorded at some fixed observation points along the expected flight paths. In these systems, it is usually assumed that the terrain is flat. Analyses are presented in this paper for selecting the observation points so that; (1) errors due to hilly terrain are minimized, (2) and all of the required scenery is available on the photographs. In addition, camera characteristics are derived so that the aerial photographs have the minimum required resolution.

Digital Photogrammetry is a recently developed area in which classical photogrammetric algorithms are applied to digitized high resolution imagery instead of using film mensuration. The major advantages of this technique are flexibility in such staple photogrammetric operations such as pass and control point selection, data reduction and the ability to enhance images digitally before measurements are made. Disadvantages include large amounts of storage required to perform many of the photogrammetric tasks and image display and review capabilities required to perform these tasks - all in a cost effective manner. In this paper we describe the architecture of a system with all the attributes of speed, storage, cost and flexibility. The characteristics of this advanced data and picture transformation system (ADAPTS) that are especially tuned to the digital photogrammetry will be highlighted. Examples of at least three photogrammetric applications will be provided. The role of these digital photogrammetric techniques in visual system data base development (a unique non-cartographic application) will be described.

In the struggle to cope simultaneously with the volume of data in an image and the display of it at real-time rates, image displays over the past fifteen years have been designed with fixed 'channels' of refresh memory, that are accessed in a raster address sequence, pixel synchronous with refresh of the CRT screen. There are significant advantages to this approach: the memory can be constructed from inexpensive mainframe computer DRAM, suitably multiplexed to achieve the desired data rate, and the addressing and control sequencing is so regular that it can be implemented with simple counters and comparators.

The integration of image processing systems with the Sun III Workstation is discussed to show the thought process used to arrive at a system integration architecture. Hardware and software design considerations are reviewed and communication protocols are examined. An example, the integration of a Viewgraphics 1024x1024 pixel 60 Hz color frame buffer, is described.

There is a historical problem with flexible, responsive real-time displays for images. The limited number of pixels displayed on the screen (usually misstated as a lack of resolution) has limited the instantaneous area of coverage of the display, while limitations in the speed of loading the display refresh memory, in conjunction with architectural and cost constraints for rapidly changing the true resolution of displayed data, have caused many users to continue a preference for film over digital displays. The key missing factor in existing displays is the ability to easily exchange area of coverage and resolution.

Commercially available broadcast video components can offer a high level of image handling and processing performance when used in traditional image processing applications. Solutions to problems of acquisition, storage, processing and display can be found by ex-ploring the wide range of equipment that is used in broadcast television. There is a great deal of similarity between many image processing requirements and the needs of a contemporary television production studio.

A perspective of a few challenging problems in the digital processing of low resolution halftone images is presented. The topics covered include screen selection, tonal reproduction control, and the scaling and resolution conversion of halftone images. Par-ticular emphasis is placed on a new method of digital scaling and resolution conversion of electronically stored, low-resolution halftone images. Many simulation examples are given.

There is increasing interest in providing office systems with the ability to digitize images of original business documents. The overall intent is to somehow integrate the disparate worlds of 'paper based' and 'electronic' information management. To fully provide this integration it will be necessary to accomplish more than simply the digitization of documents and subsequent 'pixel processing' of the bit maps. It will be necessary to effectively reverse the process of electronic document photocomposition. This will require pattern processing of the bit maps to create the same data types used in electronic creation of documents.

This paper highlights some of the hardware and software architectural issues that are pertinent to image processing systems design and operation. Any such system should also be capable of handling graphics operations as well. While no single architecture can perform optimally (in terms of speed, cost, and flexibility) in all situations (applications) it is important to note that paying attention to these issues would minimize costly design changes at a later stage. In light of the above, the paper discusses an imaging system configuration that would also support graphics and text oriented applications, and concludes with a discussion of the services and functions that should be made available on any system to support a wide variety of imaging and graphics applications.