topics in computer architecture as they relate to high performance computing. Aftel: a wide-ranging discussion of the computational characteristics and requirements of the grand challenge applications, four major architectural challenges were identified as crucial to advancing the state of the art of high performance computation in the coming decade. These computer architecture grand challenges are summarized below. PL parallel computer model provides the interface between parallel hardware and parallel software. It is the idealization of computation that computer architects strive to support with the greatest possible performance. Although a singlc: model may not fulfill the requirements of all effective architectures and applicatiion domains, the nlultitude of alternatives must be reduced to a small number to support portability of programs and reusability of program parts. ?'his challenge addresses the need for usable computer performance orders of magnitude greater than both the giga-ops performance available toda:y and the tera-ops performance which may be achieved soon. This computer perfonmance cannot be obtained by simply interconnecting massive quantities of existing prlocessor, memory, and U 0 resources. Such a system would be unmanageable to program and would ineffectively utilize its processors. The challenge is to (1) dramatically improve and (:2) effectively harness the base technologies impacting processors, memory, and UO into a computer system such that the grand challenge applications programmer has available peta-ops (10" ops) of usable processing performance. Itmerging technologies are providing an opportunity to support startling new communication-intensive applications, such as digital video workstations that treat images as easily as characters are treated today. How can computelr architecture and n.ew communications technology evolve to enable such applications? Challenge 4: Infrastructure for Prototyping Architectures Testing a new idea in computer architecture has been a difficult proce:ss requiring large investments in building design tools and providing a suitable softwan: environment for an experimental machine. Prototype development involves not on~ly hardware, but also software in the form of compilers and operating systems. A ~ I infrastructure is needed to facilitate the study of the effects of new hardware technologies and machine organizations against different application requirements. These grand challenges in computer architecture are inherently multidisciplinary and will require tea~m efforts crossing boundaries from software to hardware to applications. While it is crucial that the above challenges be addressed, it is important to stress that the viability and usability of parallel computers is also a function of the supporting software systems. Thus, a substantial effort must …

Wavelets provide a multiresolution description of images according to a wellchosen division of the space-frequency plane. This description provides information about various features present in the images that can be utilized to perform registration of remotely sensed images. In the last few years, many wavelet filters have been proposed for applications such as compression; in this chapter, we review the general principle of wavelet decomposition and the many filters that have been proposed for wavelet transforms, as they apply to image registration. In particular, we consider orthogonal wavelets, spline wavelets, and two pyramids obtained from a steerable decomposition. These different filters are studied and compared using synthetic datasets generated from a Landsat-Thematic Mapper (TM) scene.

Correlation is an extremely powerful technique for finding similarities between two images. This chapter describes why correlation has proved to be a valuable tool, how to implement correlation to achieve extremely high performance processing, and indicates the limits of correlation so that it can be used where it is appropriate. Section 4.1 gives the underlying theory for fast correlation, which is the well-known convolution theorem. It is this theory that gives correlation a huge processing advantage in many applications. Also covered is normalized correlation, which is a form of correlation that allows images to be matched in spite of differences in the images due to uniform changes of intensity. Section 4.2 treats the practical implementation of correlation, including the use of masks to eliminate irrelevant or obscured portions of images. The implementations described in this section treat images that differ only by translation, and otherwise have the same orientation and scale. Section 4.3 introduces an extension of the basic algorithm to allow for small differences of scale and orientation as well as translation. Section 4.4 presents very high precision registration. This section shows how to make use of Fourier phase to determine translational differences down to a fewhundredths of a pixel. The images must be oriented and scaled identically and have a translational difference that does not exceed half of a pixel. Section 4.5 dealswith fast rotational registration that uses phase correlation to discover the rotational difference between two images. The two images need not be scaled identically nor registered with respect to translation, but the images must have substantial overlap in order to be registered successfully. The same technique also produces an estimate of the scale difference between two images.

Registration of multiple source imagery is one of the most important issues when dealing with Earth science remote sensing data where information from multiple sensors exhibiting various resolutions must be integrated. Issues ranging from different sensor geometries, different spectral responses, to various illumination conditions, various seasons and various amounts of noise, need to be dealt with when designing a new image registration algorithm. This chapter represents a first attempt at characterizing a framework that addresses these issues, in which possible choices for the three components of any registration algorithm are validated and combined to provide different registration algorithms. A few of these algorithms were tested on three different types of datasets - synthetic, multitemporal and multispectral. This chapter presents the results of these experiments and introduces a prototype registration toolbox.

This paper shows that members of the fourier transform family are the only linear transforms that have a convolution theorem, that is, that can replace O(N 2) operations of a convolution in a time domain by O(N) operations in a transform domain. Generally, there is an additional cost to compute the transform itself. Our observation is motivated by recent activity in wavelet and subband decompositions and related spectral analyses, which are attractive alternatives for signal compression applications. A natural question when using such techniques is to determine if convolutions of N-point signals can be calculated with fewer operations in a compressed transform domain than in an uncompressed time domain. The answer is negative for a broad set of assumptions. This paper indicates what assumptions must be relaxed in seeking a linear transform that has a convolution theorem comparable to the convolution theorem for fourier transforms.

In the design of synchronous sequential machines, various canonical realizations which make use of feedback shift registers have been developed [1], [2]. This paper attempts to extend some of these results to realize asynchronous machines in a similar manner. In an earlier paper by Brzozowski and Singh [3], in which canonical feedback-free realizations of definite machines' were considered, the asynchronous unit delay (AUD) was introduced. The AUD is an n-input n-output asynchronous sequential circuit in which the present value of the output n-tuple is equal to the value of the input n-tuple before the last input change. Extending the results of the earlier paper [3] it is shown that any fundamental mode asynchronous sequential machine M can be realized by a circuit of feedback index m with one (n + m) x (n + m) AUD element and m inertial delays2 where n is the number of binary inputs and mn is the smallest integer not less than log2 S& where & is the maximum number of stable states in any column of M. Realizations with the same feedback index have been obtained elsewhere [4], [5] and the relative virtues of these different realizations are debatable. It is also shown that any fundamental mode asynchronous sequential machine M can be realized by an asynchronous circuit of feedback index 1 with 1 inertial delay and a chain of k(n + 1) x (n + 1) AUD elements. In effect the chain of AUD elements operates as a synchronous shift register. The inertial delay elements are needed in order to prevent the circuit from malfunctioning due to races and hazards. The paper is generally well written and has several examples which make it easy to read. As with the related material on synchronous realizations, the results seem to be primarily of theoretical rather than practical significance at this time. It should be noted that the realization of the AUD from basic gate elements itself requires the presence of feedback. Hence the realizations considered in this paper could really best be compared with minimum feedback realizations of synchronous machines using memory elements more complex than delay elements, such as SR flip-flops, a problem which, incidentally, has not been solved.

While automatic image registration algorithms are usually being evaluated with regards to their accuracy, it is often useful to relate this accuracy to the "initial conditions", i.e., the distance between the initial navigation geolocation and the correct result. This paper describes a modular framework that was built to describe registration algorithms, and utilize this framework to attempt to classify different registration components and algorithms in terms of their responses to the initial conditions. Performances would be evaluated on synthetic data, multitemporal and multisensor data. All results of the study would be presented at the conference and would be useful for two different purposes: (1) provide automatic quality assessment of the geolocation of remote sensing data by performing interalgorithm consistency studies; and (2) be the foundations for the design of future on-board applications including planetary exploration.

To address future NASA challenges, integration of multiple source data will be a key component, and as a first step towards this goal, very accurate registration of multi-sensor data is the first requirement for such integration. While navigation often refers to “systematic correction”, image registration refers to “precision correction.” The systematic correction is modelbased, while precision correction is feature-based. Starting from the results of the systematic correction (usually accurate within a few pixels), precision-correction utilizes selected features or control points to refine the geo-location accuracy within one pixel or a sub-pixel. Our work focuses on precision correction or automatic image registration, with the goal of achieving subpixel accuracy. We have built a modular registration framework in which different components of the registration process can be assessed and then combined in an optimal manner as a function of the application, the required accuracy and the available computational capabilities.

The study of global environmental changes involves the comparison, fusion, and integration of multiple types of remotely-sensed data at various temporal, radiometric, and spatial resolutions. Results of this integration may be utilized for global change analysis, as well as for the validation of new instruments or for new data analysis. Furthermore, future multiple satellite missions will include many different sensors carried on separate platforms, and the amount of remote sensing data to be combined is increasing tremendously. For all of these applications, the first requires step is fast and automatic registration, and as this need for automating registration techniques is being recognized, it becomes necessary to survey all the registration methods which may be applicable to Earth and space science problems and to evaluate their performances on a large variety of existing remote sensing data as well as on simulated data of soon-to-be-flown instruments. In this paper we present one of the first steps toward such as exhaustive quantitative evaluation. First, the different components of image registration algorithms are reviewed, and different choices for each of these components are described. Then, the results of the evaluation of the corresponding algorithms combing these components are described. Then, the results of the evaluation of the corresponding algorithms combining these components are presented on several datasets. The algorithms are based on gray levels or wavelet features and compute rigid transformations (including scale, rotation, and shifts). Test datasets include synthetic data as well as data acquired over several EOS Land Validation Core Sites with the IKONOS and the Landsat-7 sensors.

This paper investigates factors that degrade the precision of image registration based on phase correlation. The major sources of error are interpolation error and rotationally dependent aliasing. The latter error stems from the fact that the discrete-Fourier transform does not commute with the rotation of sampled-images, whereas in the continuous domain the corresponding operations do commute. We show through a series of examples how much the various sources of error contribute to phase-correlation registration, and we demonstrate constructive techniques for improving precision and signal to noise ratio in the registration process. Since rotationally dependent aliasing is exacerbated by the presence of high frequencies, the examples demonstrate that the use of a Blackman window removes spurious high frequencies in the spectral leakage created by the image boundary and greatly reduces aliasing effects. Since remaining aliasing effects are strongest in the low frequencies of the Fourier transform, their effects can be reduced to a negligible amount by removing frequencies within a radius of N/4 of the Fourier domain origin. A third technique is to perform phase correlation over half the Fourier plane rather than over the full plane, which more than doubles the signal-to-noise ratio of phase correlation. For an example image, the combination of techniques improved the phase-correlation signal-to-noise ratio from 8.5 to 172 and raised the peak from 0.348 to 0.885, which are substantially higher values than previously reported.

Current art uses metadata associated with satellite images to facilitate their retrieval from image repositories. Typical metadata are geographic location, time, and data type. Because the metadata do not indicate which regions within an image are obscured by clouds, retrieval with such metadata may produce an image within which the region of interest (ROI) for the user is not visible. We report a system that can automatically determine whether an ROI is visible in the image, and can incorporate this into the metadata for individual images to enhance searching capability. The goal is to annotate each image with metadata regarding a number of ROIs. An experiment with the system annotated 236 advanced very high resolution radiometer (AVHRR) images of the North Atlantic from a five-month viewing period with descriptors that expressed the visibility of an ROI centered on Long Island, NY. For ground truth, we used the classifications of three human subjects to determine visibility of the same region of interest, and labeled the ROI with the majority decision of the three subjects. Partial cloud cover made the human determination subjective, and resulted in disagreements among the subjects. Using randomly selected training subsets of the images, we found the two images whose regions were most like those in images for which the Long Island region was visible.

This paper describes a fast technique for registering image pairs from visible and infrared spectra that differ by translation, small rotations, and small changes of scale. The main result of this paper is a nonlinear prefiltering and thresholding technique that substantially enhances the cross-spectral correlation, provided that the image pairs have many features in common. We show the use of the technique in conjunction with a Fourier-based normalized correlation algorithm to perform fast cross-spectral registrations. In the absence of such prefiltering, local reversals of contrast from image to image tend to impair the quality of correlation-based registrations.The algorithm computes the translation that maximizes the overall normalized correlation of the filtered images without examining individual image features. Small rotations and scale changes can be recovered by computing the translation displacement in several different regions of the image pairs.In an experiment on a moderate-sized database, the algorithm produced a correct registration rate of over 90% at a false positive rate of less than 10%. For a particularly difficult subset of images in the database, the correct registration rate fell to approximately 85% at a false positive rate of less than 10%. This retrieval quality is comparable to that of a mutual-information-based algorithm.

This paper presents a new direct Fourier-based algorithm for performing image-to-image registration to subpixel accuracy, where the image differences are restricted to translations and uniform changes of illumination. The algorithm detects the Fourier components that have become unreliable estimators of shift due to aliasing, and removes them from the shift-estimate computation. In the presence of aliasing, the average precision of the registration is a few hundredths of a pixel. Experimental data presented here show that the new algorithm yields superior registration precision in the presence of aliasing when compared to several earlier methods and has comparable precision to the iterative method of P. Thevenaz et al. (1998).