In image database systems, we often want to retrieve images whose contents satisfy certain conditions specified in an iconic query (i.e., queries that involve input images and conditions on them). One type of image data retrieval called shape similarity-based retrieval involves retrieval of images containing one or more shapes similar to the shapes specified in the query or shapes present in the query image. In this paper, a new approach to shape similarity-based retrieval is proposed. The proposed approach is flexible enough to handle images with overlapping or touching parts. In this approach a shape is represented by a set of structural components. Each structural component is encoded as a point in multidimensional space. A similar or identical component (or shape) can be found by organizing the data in any multidimensional point access index structure. A prototype system is described in detail. Some experimental results are also presented to demonstrate the performance of the proposed technique.

We extend our previous approach to 2-D shape indexing, data-driven indexed hypotheses, to hierarchical features. It is shown mathematically and experimentally that an index based on hierarchical features is more computationally efficient than one based on nonhierarchical features. Our approach addresses two types of hierarchies: a multilevel approximation of the contours of 2-D objects and a three-level feature indexing system. Our approach can easily be extended to 3-D objects. As we know, occluded object recognition should be based on local features. However, these features may sometimes be lost due to changes of scale as well as to occlusion. Our first hierarchical mechanism is used to complement the feature loss due to scale changes. It results in multilevel approximations of the contours of objects using scale-space approaches. This approach is also beneficial when there are few boundary points of maximal curvature so that standard polygonal approximation schemes don''t work very well. Our second hierarchical mechanism is to use sets of visible local features to hypothesize the presence of objects. Verification of the various hypotheses is done via normalization and boundary template matching. Using these two hierarchical mechanisms results in an approach to shape recognition which is very general in practice. Theoretically, many types of features can be recognized, though line features are preferred. Our technique offers a new approach to the construction of indexing systems for image databases on the basis of image features.

The geometric hashing scheme proposed by Lamdan and Wolfson can be very efficient in a model-based matching system, not only in terms of the computational complexity involved, but also in terms of the simplicity of the method. In a recent paper, we discussed errors that can occur with this method due to quantization, stability, symmetry, and noise problems. These errors make the original geometric hashing technique unsuitable for use on the factory floor. Beginning with an explicit noise model, which the original Lamdan and Wolfson technique lacks, we derived an optimal approach that overcomes these problems. We showed that the results obtained with the new algorithm are clearly better than the results from the original method. This paper addresses the performance characterization of the geometric hashing technique, more specifically the affine-invariant point matching, applied to the problem of recognizing and determining the pose of sheet metal parts. The experiments indicate that with a model having 10 to 14 points, with 2 points of the model undetected and 10 extraneous points detected, and with the model points perturbed by Gaussian noise of standard deviation 3 (0.58 of range), the average amount of computation required to obtain an answer is equivalent to trying 11 of the possible three-point bases. The misdetection rate, measured by the percentage of correct bases matches that fail to verify, is 0.9. The percentage of incorrect bases that successfully produced a match that did verify (false alarm rate) is 13. And, finally, 2 of the experiments failed to find a correct match and verify it. Results for experiments with real images are also presented.

This paper describes a method for retrieving sequences of images in a motion image database using motion information as a key. This information is obtained by application of the motion analysis technique. The proposed method has three steps: automatic extraction of motion vectors through frame sequences, motion description using spatio-temporal space, and retrieval of sequences of images. The motion vectors of objects are extracted on the basis of block motion estimation. Block motion estimation tracks each block of a scene throughout a period. This tracking procedure takes account of the traceability of each block and the reliability of its calculated motion vector. Extracted motion vectors are mapped using spatio- temporal space (x-y-t). The motion of each block is represented as one vector in the feature space, and motion vectors in a scene are aggregated into several representative vectors by statistical analysis. The retrieval of sequences of images is divided into two parts: specification of the query conditions, and matching between the query conditions and the motion database. The query conditions are input from a stroke device. The user can interactively manipulate the query conditions, such as the number, period, and starting time of objects. The data are gathered as an x-y series at a given interval, and then normalized. Similarity is defined in spatio-temporal space as the distance between the query conditions and the stored motion index. Candidate sequences of images are ranked in order of distance. An experimental system has been developed for retrieving sequences of images on a workstation.

This paper describes the retrieval process from image databases, based on a description of the image content. This process is based on the assumption that a semantic interpretation of the image content is available, as the result of an automatic or semi-automatic image analysis process, applied to images belonging to specific domains and described in advance of the system. The image query language allows the user to express conditions on the symbolic representation of the images to be retrieved. It addresses the important aspects of the various image interpretations resulting from image analysis, such as partial conditions on the composition of complex objects in the image, their degree of recognition, and positional requirements on the objects in the image interpretation. The paper outlines image query processing, with particular emphasis on the access structures to support fast content-base image retrieval.

This paper presents an algorithm for pictorial information retrieval using conceptual pictorial queries. This approach is user-oriented and allows the user to construct a pictorial example as the conceptual pictorial query. It differs from other pictorial-query-by-example approaches in that the pictorial query may contain nested composite objects. Experimental results are described.

Content-based image indexing is becoming increasingly important in today''s imaging applications. Although several techniques have been developed, most have limited capabilities in dealing with relative spatial relations for multiple object images. We believe that multiple object image indexing is important for content-based indexing in many application domains. We propose a content-based indexing technique which is based on the theory of weighted center-of-mass. The index of an image is constructed by incorporating the center-of-mass of each individual object to calculate the center-of-mass of the overall image. This multiple object indexing scheme is further enhanced by incorporating existing indexing structure construction techniques, such as Jagadish''s multidimensional point representation, and Grosky''s feature index tree. An image with multiple objects can be retrieved by specifying the spatial relations among image objects, such as relative position, relative distance, and direction. The retrieval matching is based on a relative similarity measure. Furthermore, due to certain invariant properties of center-of-mass, different notions of image similarity with respect to scaling, translation, rotation, and mirrored image can be easily supported. The advantages of the proposed method are illustrated with an example of retrieving magnetic resonance images of patients'' chests within the medical image domain. The relationships to certain other indexing techniques are also addressed.

The problem of content-based retrieval of images from electronic media is especially acute in contexts where the subject-matter of the images is very general. The Birkbeck system attempts to overcome the limitations of traditional keyword systems, which provide a random and undifferentiated listing of objects in a picture, by using an augmented entity-relationship model in the data-modeling process. The interface, which allows successive query reformulation, emulates some of those aspects of browsing so important in picture collections. The subject- matter of the picture is described in terms of its objects, their attributes and the relationships existing between them. The form of the description (for both storage and retrieval) is defined by the grammar of the Picture Description Language and the Picture Query Language. Descriptions which are input to the system are parsed to generate SQL statements for the RDBMS which stores the data. Pictures relevant to the user''s enquiry are retrieved and displayed in miniature for rapid consultation. To refine further the set of pictures selected the system allows query modification, which is supported by a thesaurus. Higher rates of recall and precision are achieved than are normally possible in image storage and retrieval.

A database system is being integrated in order to store and interactively retrieve information from a several hundred Gbytes optical memory. The low cost, high reliability requirements for the development and maintenance phase of the system suggested a modular design based on a network server, optical server, and some 40 (80386 based) viewing consoles with touch screens, but no keyboard or other controls since they are exposed to use by the general public. Optical disks store graphics, video, stills, text, animation, and audio which are accessed through hypertext and interactive graphics while a somewhat simple expert analyzes and records data on various aspects of the user, such as general interests, common questions, sociological-educational, background, etc. This information is in turn used to adapt several parameters of information display: rate of flow, language style, number and type of control buttons, degree of detail, and others. The large quantity of video, still images, and different graphics formats, has made it necessary to optimize the information contained via reduction of colors/resolution, compression techniques, and recursive use of a basic set of displays and video segments. The paper presents the design in some detail, with general examples of system capabilities, growth, and applications.

The objective of this article is to offer both an image description model in three levels and an adapted image retrieval process. The proposed model allows the description of: the image globally with a classical manner (a list of concepts); the objects separately contained in the image (component objects); and relations between the component objects. This allows more semantic aspects of the image structure to be expressed. It is also possible to use fuzzy information in each level of the description. To build this model, we used the concepts of the object-oriented approach: generalization/specialization, aggregation, and inheritance. This approach is born from a strong need expressed in several fields such as medicine, CAD/CAM, architecture, and the teledetection industry, to be able to retrieve and describe images, their component objects, and the relations between them more precisely, with their own vocabulary and concepts.

This paper describes visual interaction mechanisms for image database systems. The typical mechanisms for visual interactions are query by visual example and query by subjective descriptions. The former includes a sketch retrieval function and a similarity retrieval function, while the latter includes a sense retrieval function. We adopt both an image model and a user model to interpret and operate the contents of image data from the user''s viewpoint. The image model describes the graphical features of image data, while the user model reflects the visual perception processes of the user. These models, automatically created by image analysis and statistical learning, are referred to as abstract indexes stored in relational tables. These algorithms are developed on our experimental database system, the TRADEMARK and the ART MUSEUM.

Despite their similarities, hypertext and information retrieval are generally considered as separate research areas. We aim at gathering hypertext and information retrieval into a common problem set by proving that our contribution to information retrieval involves an important overlap with hypertext ideas. Our initial field is the study of interactivity in retrieving images. It''s with this aim in view that we have designed the EXPRIM process implemented in the RIVAGE prototype as a machine-learning process by applying symbolic learning techniques.

We present a query system for an object-oriented biomedical imaging database containing 3-D anatomical structures and their corresponding 2-D images. The graphical interface facilitates the formation of spatial queries, nonspatial or symbolic queries, and combined spatial/symbolic queries. A query editor is used for the creation and manipulation of 3-D query objects as volumes, surfaces, lines, and points. Symbolic predicates are formulated through a combination of text fields and multiple choice selections. Query results, which may include images, image contents, composite objects, graphics, and alphanumeric data, are displayed in multiple views. Objects returned by the query may be selected directly within the views for further inspection or modification, or for use as query objects in subsequent queries. Our image database query system provides visual feedback and manipulation of spatial query objects, multiple views of volume data, and the ability to combine spatial and symbolic queries. The system allows for incremental enhancement of existing objects and the addition of new objects and spatial relationships. The query system is designed for databases containing symbolic and spatial data. This paper discuses its application to data acquired in biomedical 3- D image reconstruction, but it is applicable to other areas such as CAD/CAM, geographical information systems, and computer vision.

With a potential of up to 1012 bits/cm3, storage medium based on volume holographic memory can be the next generation of optical memory systems. However, any storage media for holographic memory must have large reflective index modulation, large available volume material while simultaneously preserving high response speed and low crosstalk. Photorefractive crystals, such as LiNbO3, SBN, and BaTiO3, are the most popular materials for the 3-D memory medium candidates. However, they possess low refractive index modulation, very small available volume, low response speed, and poor optical quality. They also exhibit low multiplexing capability, low diffraction efficiency, and high cross talk. These are the main difficulties for the application of photorefractive materials for holographic memory. Organic materials are of high nonlinearity, easy processing, and great cost-effectiveness. They may overcome the problems with photorefractive crystals. In this paper, an erasable dye polymer memory medium is exploited for a randomly addressable read/write/erase holographic memory system. To our knowledge, this is the first demonstration of an erasable holographic memory system. The system has 100 pages of memory. Its read/write speeds are mainly limited by the input spatial light modulator/output detector array. Although being 2-D, it can be potentially developed into a 3-D system.

Electron trapping (ET) materials are novel storage media developed by Quantex. The architecture of a three-dimensional stacked layer optical memory based on ET materials is reported in this paper. Two-dimensional page memories are individually stored in ET layers by properly imaging the page composer onto a specific layer with blue light. To read out data in a specific layer, a slice of IR beam is addressed into the ET layer from the side of the stacked layer media. The addressed ET thin film contains the IR read beam like a slab waveguide. The orange emission corresponding to written page memory at that layer is emitted as a result of the IR stimulation. System design considerations and preliminary experimental results are presented.

The diffraction efficiency and nonlinear transmission properties of chemically enhanced thin films of bacteriorhodopsin are analyzed by using absorption spectroscopy, the Kramers-Kronig transformation, coupled wave theory, and a simplified kinetic model of the bacteriorhodopsin photocycle. Photoconversion of bR to a 50:50 mixture of bR and M generates a large change in refractive index that is proportional to the bacteriorhodopsin concentration and is greatest in regions where the difference in absorption coefficients is smallest. The predicted diffraction efficiencies are dominated by large phase components in regions of minimal bR and M absorption. The maximum diffraction efficiency (11) for a 2.5 OD, 150 micrometers thick film occurs at readout wavelengths between 620 nm - 700 nm. These films also exhibit significant nonlinearity in transmissivity at low laser intensities and could find potential use in spatial filtering applications. A real time optical associative memory based on holographic thin films of bacteriorhodopsin is also discussed.

The need to store large amounts of information in small volumes cannot be overstated, nor can the need for these memories to be capable of large bandwidths and parallel access of the information be overemphasized. Three-dimensional (3-D) storage may provide a desirable solution to these needs. The various possible avenues which may lead to 3-D information storage include hole burning, phase hologram, and two photon processes using organic semiconductors and biomolecules such as bactiochadorpsin. In this paper we restrict discussion to the last topic, namely, 3-D storage by means of two-photon absorption, particularly in the utilization of organic photochromic materials. In previous communications we have presented experimental data which suggest a possibility for constructing useful 3-D volume memory devices based upon two-photon virtual excitation processes. In this paper we discuss the means for writing and reading information within a 3-D storage memory, describe the materials used, and present a status report on their relevant properties. The theoretical bases for two-photon processes were established long ago, originating theoretically in 1932 and experimentally in the early 1960s. The probability for a two-photon transition may be expressed as a function of three parameters: line profile, transition probability for all possible two-photon processes, and intensity.

We describe a new approach to develop a high-speed, high-density random access cache memory. This new technique is based on a novel approach using, in essence, a hybrid of the time domain stimulated echo (SE) concept and the frequency domain scheme. It has been demonstrated that the rare-earth doped crystals, can at low temperatures store large amounts of information for up to 24 hours using SE. The basic approach we have developed is the partitioning of the absorption frequency domain into smaller bins, so that each frequency bin stores a smaller portion of information independently. The bins are distinguishable by their different absorption frequencies and they are accessed by changing the laser frequency (color). However, information is still stored and retrieved using the time-domain pulse sequency used in SE. Specifically, with the laser frequency set at the particular absorption of a frequency bin, the laser is externally modulated by an acousto-optic modulator to produce amplitude modulated pulses representing the digital data, the write pulse, and the read pulse, during the write and retrieve operation, respectively. The advantages of this approach are that memory can be stored in space, time, and frequency domain, allowing us to tailor a flexible memory architecture to match those of the computational processor.

In order to achieve the promise of terabit/cm3 data storage capacity for volume holographic optical memory, two technological challenges must be met. Satisfactory storage materials must be developed and the input/output architectures able to match their capacity with corresponding data access rates must also be designed. To date the materials problem has received more attention than devices and architectures for access and addressing. Two philosophies of parallel data access to 3-D storage have been discussed. The bit-oriented approach, represented by recent work on two-photon memories, attempts to store bits at local sites within a volume without affecting neighboring bits. High speed acousto-optic or electro- optic scanners together with dynamically focused lenses not presently available would be required. The second philosophy is that volume optical storage is essentially holographic in nature, and that each data write or read is to be distributed throughout the material volume on the basis of angle multiplexing or other schemes consistent with the principles of holography. The requirements for free space optical interconnects for digital computers and fiber optic network switching interfaces are also closely related to this class of devices. Interconnects, beamlet generators, angle multiplexers, scanners, fiber optic switches, and dynamic lenses are all devices which may be implemented by holographic or microdiffractive devices of various kinds, which we shall refer to collectively as holographic interconnect devices. At present, holographic interconnect devices are either fixed holograms or spatial light modulators. Optically or computer generated holograms (submicron resolution, 2-D or 3-D, encoding 1013 bits, nearly 100 diffraction efficiency) can implement sophisticated mathematical design principles, but of course once fabricated they cannot be changed. Spatial light modulators offer high speed programmability but have limited resolution (512 X 512 pixels, encoding about 106 bits of data) and limited diffraction efficiency. For any application, one must choose between high diffractive performance and programmability.

High speed random access memory (RAM) components used in current memory subsystems, caches, and DMA channels as well as in super high speed computer systems, rely on expensive and high power consumptive ECL and GaAs memory chips. These memory devices provide only marginal memory densities per chip (4 - 16 k bit). It is anticipated that the integration of optical interconnect technology coupled with alternative GaAs structures, can not only increase the speed of such systems but also decrease the power consumption. A frequency tunable, multiple quantum well, distributed Bragg reflection, GaAs solid state, laser diode array is being designed in conjunction with high resolution, planar waveguide optical gratings to achieve a low inertia and, therefore, a low-power sub-nanosecond parallel (e.g., 64 bits wide), angular scanning capability. When applied to an optical memory media such as dichromated gelatin holograms or a magneto-optic material which is used in optical disk technology, this photonic random optical memory access eliminates the power consumptive address system in RAM. Ultimately, disk or memory access speed will be increased by several orders of magnitude, while keeping power consumption at a minimum.

A motionless head 2-D parallel readout system for optical disks is presented. The system is designed to read data blocks encoded as 1-D Fourier holograms distributed radially on the disk active surface. Such systems offer several advantages: high data rates, low retrieval times, and simple implementation. It is used as the secondary storage of a high performance optoelectronic associative memory system.

We have developed a document station for automated entry and remote retrieval of image documents, and we present experimental results on remote retrieval via an integrated services digital network (ISDN). The document station consist of a UNIX workstation, a 130 mm rewritable magneto-optical disk subsystem, a code/decode board, character-recognition hardware, and a high-speed image scanner. This station can store 300,000 black-and-white A4 pages at 5.0 seconds per page. The experimentally measured time for retrieving one page at a remote terminal is 1.5 seconds within a 10 Mbps local-area network (LAN) and 2.0 seconds within a 1.5 Mbps ISDN.

The coding, storage, and reconstruction of images is a major concern in the application of computer technology to technical and scientific problems. In such applications it is highly desirable to reduce the storage and transmission requirements for image data. An image can be coded compactly when it is possible to exploit self similar redundancy in the image. Our research has focused on the development of a `fractal'' method for compressing image data. Our approach to image compression, similar to Jacquin, is to tessellate the image with a tiling which varies with the local image complexity, and to check for self similarity among the tiles. Self similarities are coded as systems of affine transformations which can be stored far more compactly than the original image. This method is inherently lossy, since the self similarities are never exact. We have tested our encoding scheme on a variety of test images, gaining good compression ratios. At high compression ratios, the scheme yields better signal to noise ratios than are reported for other techniques. Our scheme is versatile in that it allows a trade- off between compression, reconstructed image fidelity and encoding time. Our methods are computationally intensive but are feasible for non-real time applications on workstations or main frame computers. We are currently studying applications such as multimedia systems and CD-ROM mass storage systems. The algorithms can be accelerated considerably by dedicated hardware for real time requirements. Fractal compression is a promising approach to image compression. Within a very short development time, fractal techniques have yielded results which rival the best examples of data compression afforded by other methods. Although fractal encoding of images is complex and may require specialized hardware for real time applications, the decoding process can be widely utilized because it is simple, fast, and suitable for software implementation. Thus, it can be run on workstations or personal computers without special requirements. This report contains a description and motivation of the encoding algorithm, a description of a new decoding algorithm, results, comparison with results in the literature, and a discussion of generalizations to the algorithm.

Optical disk storage systems are finding greater application within the U.S. Air Force. A range of commercial, ruggedized, and militarized systems have been fielded for user test and evaluation under realistic operating conditions. The success of these initial systems promises to foster greater user confidence and willingness to explore new applications. Within the Department of Defense, Rome Laboratory is a leader in developing high-performance optical disk systems for military applications.

There is a need for a cost-effective approach to training ground forces in combat vehicle identification (CVI), especially under thermal sighting conditions. Computer image generation (CIG) provides many advantages for the development of cost effective CVI training: 3-D and thermal modeling of vehicles, thermal rendering of landscape scenes, and a wide assortment of training scenarios. But the extent of computation required for CIG makes it impossible to combine real-time presentation and interactive control, except on very expensive, high-end graphics systems. We describe a solution in which we have separated CIG and presentation, allowing the computation-intensive processes to run `off-line'' on virtually any computer (high- end graphics or general-purpose). The presentation is provided by digital video interactive (DVI) technology, a flexible, low-cost, PC-based system for delivering high quality graphics, photo quality imagery, motion video, and audio. We have developed and demonstrated a specialized animation algorithm which combines landscape scenes and vehicle imagery to produce animated scenarios in which trainees face varying degrees of difficulty in vehicle identification. The simulation is intended to replicate the appearance and controls of common thermal sights in tanks. The animation has been integrated with a complete training program, which includes guided instruction in vehicle recognition.

Synthetic aperture radar generates large images. The storage and retrieval of these images from an archive supports the operational utilization of newly generated images. In addition to archiving the new images for reference, supporting images are stored in the archive. An optical disk jukebox with a capacity of 500 gigabytes to 1 terabyte provides an essential element of the storage of the archived images. The optical jukebox supports the rapid retrieval of images that are as large as 20 megabytes with no compression. With compression, the images that can be effectively stored on the optical jukebox can be as large as 200 megabytes. Images are typically pre-fetched a few minutes before they are needed. Ad-hoc images may be fetched in less than a minute for the smaller images. For the larger images the time to fetch the image is established by the time to read the image from the optical disk. This paper compares the storage on the optical jukebox with competing storage on helical scan tape.

Write-Once optical disk technology is an application-specific medium that offers excellent reliability and security while storing massive amounts of data. The storage capability is critical in that a single disk, or platter, can store as much as 6.55 gigabytes of information which cannot be altered or erased. However, the retrieval process is also significant since optical disks allow for non-linear, random access to all data. Although Write-Once optical disk system applications are very diverse, they all express a common denominator -- the need to store vast amounts of data and the ability to quickly and effortlessly retrieve the information. Write-Once optical disk technology has rapidly matured in vertical markets, resulting in increased productivity, space conservation, and cost savings. Applications ranging from mass storage and retrieval of legal documents to mug shots, fingerprints, and x rays illustrate how Write- Once optical disk technology has become a strategic technological tool in vertical markets. The applications also demonstrate how the technology has expanded its functionality beyond routine storage and retrieval.

This paper provides an update on the development status of two high performance optical disk drive systems being developed by GE Aerospace for the US Government. One is for airborne platforms the other is for long life spaceflight platforms. Device data rates are 3 and 37 MBytes/sec sustained capacities are 12 and 10 GBytes and access times are 350 and 100 ms respectively. Flight tests for each are scheduled for 1993 and 1995. 1.

We are building a networked multimedia environment called Continuous Network Media Server which is based on the client-server model. The purpose of the server is to provide users with a uniform environment of multimedia where users can employ each medium or any combination of media. We have developed an experimental system of audio server to investigate problems in handung continuous media on an ordinary workstation. The experimental results show the effectiveness of our approach and highlight several problems related to current workstation technology. 1

The economics of large electronic libraries requires a hierarchy of storage devices, where the lowest level contains the most data and has the poorest performance characteristics. In today''s electronic library systems, this is the role of the optical jukebox. Performance obtainable from an electronic library is a function of many system and application characteristics. We are addressing performance of client/server libraries where retrieval is the primary interaction of users with the library. When the documents to be retrieved reside on the optical media managed by the jukebox server, and nowhere else in the hierarchy, then the retrieval of these documents becomes a batch task. This is acceptable for many users and applications environments: for professionals who are simultaneously at various stages of many tasks and can anticipate the next ones, and for clerical workers in an environment controlled by work- flow (and not interrupt driven). Scheduling of batch retrievals of documents from archival storage need not be in the same order as requests. This paper proposes a cyclic scheduling for retrieval of documents from the optical jukebox, and shows by simulation that this cyclic scheduling can offer significant performance improvement over that obtained by FCFS scheduling of the jukebox.

In this paper we describe a technique for performing shaped-based content retrieval of images from a large database. In order to be able to formulate such user-generated queries about visual objects, we have developed an hierarchical classification technique. This hierarchical classification technique enables similarity matching between objects, with the position in the hierarchy signifying the level of generality to be used in the query. The classification technique is unsupervised, robust, and general; it can be applied to any suitable parameter set. To establish the potential of this classifier for aiding visual querying, we have applied it to the classification of the 2-D outlines of leaves.