It is a challenging issue to provide a user-friendly means to retrieve information from a very large database when the user cannot clearly define what the information must be. To find images which are most relevant to the given description, we proposed a Fuzzy Interactive Activation and Competitive (FIAC) neural network model. There are two layers of units in a FIAC: concept layer and example layer. The example layer stores data and feature measures of images while the concept layer represents the attributes of the images. The links between the concept units and the example units are fully connected and bidirectional-directional with different link function for each direction. To ensure best matches, inhibitory links are created among example units and among conceptual units in the same pool. When a query is defined by specifying truth values for a set of concept units, the neural network will run for a given number of cycles and the images relevant to those fuzzy subsets will have the maximum activation. The FIAC neural net for retrieval has been tested for face images, and produced very promising results.

The number of nodes of an edge quadtree representing a figure is the measure of its space complexity. This number depends on the figure's shape, its resolution and its precision. The goal of this work is to find a function which yields the number of nodes of an edge-quadtree when these three parameters are specified in input. A unique value to represent both the resolution and the precision is used. To measure the shape of the image we use the fractal dimension and a methodology to calculate the fractal dimension and the fractal measure in the case of a discrete image is proposed. Given these three parameters we use a neural network to approximate the function. The computational results show the effectiveness of this approach.

The fuzzy frame approach to representation for images and image features is offered. The essence of the given approach consists that database represents a semantic network of frames with fuzzy fillings for slots (fuzzy frames) and the query processing is a decision making in fuzzy situation. The decision making process is formalized as the search of the best path in the frame tree. This approach can be the methodological basis for development of intelligent knowledge bases under the images, since the semantic filling of frame can be most different (purely images, textual records, lists of characteristic or correspondent points, lists of objects and etc.).

With the advances in imaging and storage technology, critical document handling and archiving applications have been migrating to digital technology, thereby reducing the cost and improving the quality and services. Documents are first sorted, scanned and compressed into digital images. Then state-of-the-art computer and networking technology is used to manage these images. In this paper we discuss the architecture and performance of a large, geographically distributed archival and retrieval system for image and document applications. The architecture of interest is a hierarchical one. We examine performance characteristics of the architecture and identify critical resources and parameters of system performance. Using decomposition and aggregation techniques, we model subsystems and study the end-to-end system performance. We show the performance/cost tradeoffs and various approaches to further optimize the system performance.

Recent attention on global environmental changes has stimulated the development of large scale global information systems. Satellite images play a very important role for understanding these global changes, but their data size are very large. Magnetic tape archivers are often used for them because of their capacity, but their capacity cannot be increased except by adding a new archiver which is independent of the others. We designed a scalable tape archiver, which consists of several element archivers and can be extended to any number of archivers. Each element archiver can transfer a cassette to a neighboring one. In the scalable tape archiver, performance strongly depends on data placement and usage of the tape drives. In this paper, we propose several cassette tape migration algorithms to balance the load across the element archivers and also evaluate the performance of the various proposed algorithms through simulation.

Medical radiographs and associated data collected as part of a nationwide health survey in the U.S. are digitized and stored in an electronic archive accessible over the Internet. This paper describes the prototype system developed for the archiving of the data and the client software to enable a broad range of end users to access the archive, retrieve text and image data, display the data and manipulate the images.

The Data Services Segment of the Defense Mapping Agency's Digital Production System provides a digital archive of imagery source data for use by DMA's cartographic user's. This system was developed in the mid-1980's and is currently undergoing modernization. This paper addresses the modernization of the imagery buffer function that was performed by custom hardware in the baseline system and is being replaced by a RAID Server based on commercial off the shelf (COTS) hardware. The paper briefly describes the baseline DMA image system and the modernization program, that is currently under way. Throughput benchmark measurements were made to make design configuration decisions for a commercial off the shelf (COTS) RAID Server to perform as system image buffer. The test program began with performance measurements of the RAID read and write operations between the RAID arrays and the server CPU for RAID levels 0, 5 and 0+1. Interface throughput measurements were made for the HiPPI interface between the RAID Server and the image archive and processing system as well as the client side interface between a custom interface board that provides the interface between the internal bus of the RAID Server and the Input- Output Processor (IOP) external wideband network currently in place in the DMA system to service client workstations. End to end measurements were taken from the HiPPI interface through the RAID write and read operations to the IOP output interface.

This paper examines the feasibility of networked environments as the basis for image information systems. Experiments are conducted to investigate the effect of the network on an image display application, using an Ethernet network. It is shown that for typical configurations the network is not the major bottleneck in display of large images. The major factors are the display software used, CPU performance and the workstation memory capacity. To generate significant delays due to network traffic with a small workgroup setup, an artificial configuration was required that would not be employed in practice. To support larger image datasets and greater numbers of users, options for enhancing the network are examined.

Image archival and retrieval makes significant demands on compression technology. An image library will be large and heterogeneous--thus the need for deep lossy compression and robust performance across diverse content. For long term archival, the codec must provide a (near) lossless mode. Ideally, one file format will support both. It should allow progressive transmission and be multi-resolution, supporting full-size, screen-size, and thumbnail-size versions. This paper presents a new spline-based image compression scheme that satisfies these objectives.

A progressive region-based image compression technique is presented. This technique is based on the second generation image compression paradigm, representing images as a union of homogeneous regions in harmony with the characteristics of the human visual system. The technique is progressive, allowing for an initial coarse version of the image to be reconstructed quickly from the compressed image data, and then refined through to the desired degree of fidelity. This paper describes how this technique can be applied to human face image archiving and retrieval. Compression results are provided and comparison with the JPEG standard is undertaken.

The progressive transmission of digital images is an important field of interactive image communications over low-bandwidth channels. An effective progressive transmission system should be able to build up images with highest possible fidelity at every stage of transmission, so that a viewer can respond to the received image as quickly as possible. In this paper, a perceptually tuned progressive image transmission scheme based on removing just-noticeable distortion (JND) from the image in pyramid form is presented. The JND profile is derived from the analysis of local properties of image signals. According to the sensitivity of human visual perception to spatial frequency, a distortion allocation algorithm is developed to find the associated JND profile of each pyramid level to sift out perceptually significant signals and determining their quantizer stepsizes. Simulation results show that images reconstructed by the proposed coding schemes can be perceptually more pleasing than those reconstructed by the JPEG progressive decoder at earlier stages of transmission.

Applications using remote sensing imagery rely on very high accuracy of the pixel data values. This requirement necessitates use of lossless or near-lossless compression techniques during storage. These techniques tend to provide rather limited compression rates of between 2:1 and 3:1. An approach for improving the compression rates of lossless mode JPEG based on contextual choice of the current pixel predictor is presented. Tests on a range of remote sensing images demonstrate the promise in this approach.

In this paper we present a progressive template matching algorithm that can be used when performing content-based retrieval on images or videos that are stored using DCT based block transforms such as those used in the JPEG compression standard. In the proposed method, the template matching is initially performed on a low-resolution version of the image consisting of the low frequency coefficients from the DCT-transformed image. The template is then matched against the neighborhood of the resulting hit(s) by incorporating additional DCT coefficients into the analysis. We have conducted preliminary experiments on a database consisting of large satellite images; our results show that the progressive template matching methodology yields significant computational speedup over the conventional approach.

Lines and junctions are principal features for a line image. The spatial relationships established among them are usually employed in applications such as OCR and architectural blueprint vectorization. The conventional thinning techniques often suffered the pitfall of spurious junction points that are crucial features to derive. The liability of the shape of the skeleton resulting from multiple fork points connected through several short branches will impeded the further recognition stage. In this paper, a new approach which differentiates from the conventional thinning algorithm in vectorizing a raster line image is presented. This vectorization algorithm takes the ensemble of pixels within the line segments collectively as legitimate candidates in deciding the vectorized representation. This method can not only segment the lines and junctions but also construct their spatial relationships. A maximal inscribing circle (MIC) concept is introduced to derive the directions of line segments. An iterative procedure is developed to identify each line segment and the corresponding junctions. Experimental studies comparing the performance of a conventional thinning method with that of our MIC algorithm are performed using the flow diagram and logical diagram as test images. The results demonstrate that our approach is computation efficient, robust and may render optimal multi-pixel-width vectorized line representation at user's discretion.

The Radiologic Image Communication and Archive (RICA) service is designed to provide a shared archive for medical images to the widest possible audience of customers. Images are acquired from a number of different modalities, each available from many different vendors. Images are acquired digitally from those modalities which support direct digital output and by digitizing films for projection x-ray exams. The RICA Central Archive receives standard DICOM 3.0 messages and data streams from the medical imaging devices at customer institutions over the public telecommunication network. RICA represents a completely scalable resource. The user pays only for what he is using today with the full assurance that as the volume of image data that he wishes to send to the archive increases, the capacity will be there to accept it. To provide this seamless scalability imposes several requirements on the RICA architecture: (1) RICA must support the full array of transport services. (2) The Archive Interface must scale cost-effectively to support local networks that range from the very small (one x-ray digitizer in a medical clinic) to the very large and complex (a large hospital with several CTs, MRs, Nuclear medicine devices, ultrasound machines, CRs, and x-ray digitizers). (3) The Archive Server must scale cost-effectively to support rapidly increasing demands for service providing storage for and access to millions of patients and hundreds of millions of images. The architecture must support the incorporation of improved technology as it becomes available to maintain performance and remain cost-effective as demand rises.

As our applications continue to become more sophisticated, the demand for more storage continues to rise. Hence many businesses are looking toward data warehousing technology to satisfy their storage needs. A warehouse is different from a conventional database and hence deserves a different approach while storing data that might be retrieved at a later point in time. In this paper we look at the problem of storing and retrieving medical image data from a warehouse. We regard the warehouse as a pyramid with fast storage devices at the top and slower storage devices at the bottom. Our approach is to store the most needed information abstract at the top of the pyramid and more detailed and storage consuming data toward the end of the pyramid. This information is linked for browsing purposes. In a similar fashion, during the retrieval of data, the user is given a sample representation with browse option of the detailed data and, as required, more and more details are made available.

A photograph imaging system presents a unique set of requirements for indexing and retrieving images, unlike a standard imaging system for written documents. This paper presents the requirements, technical design, and development results for a hierarchical ANSI standard thesaurus embedded into a photograph archival system. The thesaurus design incorporates storage reduction techniques, permits fast searches, and contains flexible indexing methods. It can be extended to many applications other than the retrieval of photographs. When photographic images are indexed into an electronic system, they are subject to a variety of indexing problems based on what the indexer `sees.' For instance, the indexer may categorize an image as a boat when others might refer to it as a ship, sailboat, or raft. The thesaurus will allow a user to locate images containing any synonym for boat, regardless of how the image was actually indexed. In addition to indexing problems, photos may need to be retrieved based on a broad category, for instance, flowers. The thesaurus allows a search for `flowers' to locate all images containing a rose, hibiscus, or daisy, yet still allow a specific search for an image containing only a rose. The technical design and method of implementation for such a thesaurus is presented. The thesaurus is implemented using an SQL relational data base management system that supports blobs, binary large objects. The design incorporates unique compression methods for storing the thesaurus words. Words are indexed to photographs using the compressed word and allow for very rapid searches, eliminating lengthy string matches.

In this paper, we develop an indexing scheme for medical images. In general, for a given medical image, there are objects which are clinically important amongst the rest. We name the objects as the dominant objects. Our proposed index is composed of three parts: (1) dominant objects in images are located; (2) each image will have an associated R-tree which is constructed by its dominant objects; and (3) an R-tree that clusters similar images together. To demonstrate the effectiveness of the index developed, we use images of skin lesions as the image data. Our initial experiments give promising results for image retrieval.

This paper presents an experimental evaluation of different image content representations, all of which are based on the use of color histograms to support indexing and searching schemes. We investigate the use of different color resolutions, restriction to dominant colors, and matching based on both global and local histograms. We also examine how suitable numerical index keys may be designed to support retrieval, and we assess the use of Self-Organizing Maps to guide structuring a database of images. All of our results are based on experimental studies, and our conclusions should lead to useful guidelines for developing image indexing and retrieval systems based on visual content.

Linear predictive techniques perform poorly when used with color-mapped images where pixel values represent indices that point to color values in a look-up table. Re-ordering the color table, however, can lead to a lower entropy of prediction errors. In this paper we investigate the problem of ordering the color table such that the absolute sum of prediction errors is minimized. The problem turns out to be intractable, even for the simple case of 1D prediction schemes. We given two heuristic solutions for the problem and use them for ordering the color table prior to encoding the image by lossless DPCM like techniques. The first heuristic is based on a simulated annealing approach and is computationally expensive. The second heuristic, however, is simple and sacrifices optimality for computational efficiency. It involves successive merging of ordered sets of color table entries until all the entries have been merged into a single set. Simulation results giving comparison of the two heuristics with previous approaches are presented. It is seen that significant improvements can be obtained with the proposed heuristics. We then use a simple error modeling technique to encode prediction residuals and demonstrate the improvements in actual bit rates that can be achieved over dictionary based coding schemes that are commonly employed for color-mapped images.

A method based on the joint indexing of spatial and spectral information is presented for the purposes of content-based color and multispectral image retrieval. The image representation consists of hierarchical spatial structuring along with feature extraction in localized regions of the image. The algorithm comprises of a quadtree based image splitting method and a clustering technique to extract spectral information. It can support localized queries regarding either the colors or classes present in color and multispectral images. Experimental results suggest that the joint spatial and spectral indexing approach is a very flexible and efficient method for content-based queries in image database management.

With advances in storage, processing and communications technologies the creation of large collections of image data has become feasible, for which Picture Retrieval Systems will require flexible and efficient retrieval methods. The way a user searches for an object such as an image will depend on the object's context and the application in mind. In this paper we review methods of indexing imagery based on information associated with, and information derived from images. Experimental systems are described which illustrate different methods for indexing according to imagery type. One system indexes a collection of ship images. Since the focus is an object, namely a ship of interest, the indexing method adopted is one that can describe relative relationships of components of the ship. The second system is designed to manage collections of images of geographic areas, such as aerial or satellite image. In this case the images are indexed by geographic locations using a geographic information system to manage the spatial data structures. An extension of this system is described which allows the capture of object-related information. This allows object based descriptions to be incorporated for retrieval. Implementation of each system is described and example output presented.

Evaluation is a critical issue in any information systems. This problem has become more and more important with the rapid development of multimedia systems. Feature measures and similarity measures play a central role in content-based retrieval. Evaluation of their effectiveness and efficiency then become a key issue in assessing the performance of a content- based multimedia system. A learning algorithm has been studied to find a suitable and hopefully the best similarity function for a given set of feature measure and a given set of training data set.

In this paper, we propose a new technique based on wavelet vector quantization for the storage and retrieval of compressed images. Here, the images are first decomposed using wavelet transform followed by vector quantization of the transform coefficients. We note that similar images map to similar labels. Hence, the labels corresponding to an image constitute a feature vector which is used as an index to store and retrieve the image. In addition, the lowest resolution subimages resulting from the wavelet decomposition serve as visual icons for browsing purposes. The proposed technique provide fast access to the compressed images in the database has a lower cost for computing and storing the indices compared to other techniques reported in the literature.

Image databases typically manage feature data that can be viewed as points in a feature space. Some features, however, can be better expressed as a collection of points or described by a probability distribution function (PDF) rather than as a single point. In earlier work we introduced a similarity measure and a method for indexing and searching the PDF descriptions of these items that guarantees an answer equivalent to sequential search. Unfortunately, certain properties of the data can restrict the efficiency of that method. In this paper we extend that work and examine trade-offs between efficiency and answer quality or effectiveness. These trade-offs reduce the amount of work required during a search by reducing the number of undesired items fetched without excluding an excessive number of the desired ones.

We propose a new method for indexing large image databases. The method incorporates neural network learning algorithms and pattern recognition techniques to construct an image pattern dictionary. Image retrieval is then formulated as a process of dictionary search to compute the best matching codeword, which in turn indexes into the database items. Experimental results are presented.

A novel and efficient pose-invariant guided template matching algorithm, for object recognition in images, is proposed. Template matching is performed in a rotation-scale- translation (pose) invariant fashion, thus greatly reducing the 4-d search space to a single point. The invariance is achieved by preprocessing the input image and associating certain geometric descriptors with the objects in the image. These descriptors completely characterize the affine parameters associated with the objects which must be applied to candidate templates, and the location in the image where the template is to be applied for a match. Preprocessing and matching are performed on a wavelet pyramidal decomposition of the image in a multiresolutional coarse-to-fine fashion for computational efficiency. An efficient search strategy is also proposed for selecting templates in the template database.

We present two new approaches based on color histogram indexing for content-based retrieval of image databases. Since the high computational complexity has been one of the main barriers towards the use of similarity measures such as histogram intersection in large databases, we propose a hierarchical indexing scheme where computationally efficient features are used to subset the image before more sophisticated techniques are applied for precise retrieval. The use of histograms at different color resolutions as filtering and matching features in a hierarchical scheme is studied. In the second approach, a multiresolution representation of the histogram using the indices and signs of its largest wavelet coefficients is examined. Excellent results have been observed using the latter method.

This paper suggests a wavelet transform based multiresolution approach as a viable solution to the problems of storage, retrieval and browsing in a large image database. We also investigate the performance of an optimal uniform mean square quantizer in representing all transform coefficients to ensure that the disk space necessary for storing a multiresolution representation does not exceed that of the original image. In addition, popular wavelet filters are compared with respect to their reconstruction performance and computational complexity. We conclude that, for our application, the Haar wavelet filters offer an appropriate compromise between reconstruction performance and computational efforts.

This paper presents an approach to texture-based image retrieval which determines image similarity on the basis of the matching of fractal codes. Image fractal codes are generated via a fractal image compression technique that has been recently proposed as an effective image compression method. Each image is represented by a set of self-transformations through which an approximation of the original image can be reconstructed. These self-transformations, which are unique to each image and are semantically rich, are termed fractal codes. An image data model is proposed which constructs each image as a hierarchical structure. Each image is decomposed into block-based segments which are then assembled by a hierarchy on the basis of inclusion relationships. Each segment is then fractally encoded. The fractal codes of an iconic image are used as texture key and are matched with the fractal codes of images in a database by applying searching and matching algorithms to the hierarchies of the database images to locate the segments which best match the fractal codes of the iconic image. Retrievals of both exact and inexact matching of images are supported.

In this paper, we describe a prototype system, named DAIRS, for distributed image retrieval. DAIRS features adaptive query reformulation mechanism for improving the retrieval effectiveness. The query reformulation mechanism is based on the calculation of the functional dependency between each image attribute and the user's relevance feedback. The importance (or weight) of each attribute is modified in the reformulated query based on the degree of such functional dependency. Since image servers are dynamically evolving in a distributed environment, DAIRS has been designed to deal with image databases in various domains distributed in the Internet. The DAIRS communication protocol is designed to cooperate with http and ftp, so that the client can easily access the distributed image repositories. Experimental results show that the query reformulation mechanism significantly improves the retrieval effectiveness.

The problem of scenic image classification is presented in the paper. On considering the specific nature of this problem, we propose a statistically data-based method, the Hidden Markov Model, to solve this problem. We segment an image and use the sequence of segments as the definition of the image; we then train a HMM on a test set of sequences/images to establish a classification. We present preliminary results on the use of a 1D HMM for classification of images as either indoor or outdoor.

Thousands of documents and images are generated daily both on and off line on the information superhighway and other media. Storage technology has improved rapidly to handle these data but indexing this information is becoming very costly. HNC Software Inc. has developed a technology for automatic indexing and retrieval of free text and images. This technique is demonstrated and is based on the concept of `context vectors' which encode a succinct representation of the associated text and features of sub-image. In this paper, we will describe the Automated Librarian System which was designed for free text indexing and the Image Content Addressable Retrieval System (ICARS) which extends the technique from the text domain into the image domain. Both systems have the ability to automatically assign indices for a new document and/or image based on the content similarities in the database. ICARS also has the capability to retrieve images based on similarity of content using index terms, text description, and user-generated images as a query without performing segmentation or object recognition.

The visual information browsing environment, VIBE, is one of a series of visual information retrieval interfaces being developed, with the aim of permitting a user to access a database dynamically using several simultaneous approaches. Thus VIBE permits a carefully tailored, highly individual approach to information organization and retrieval. We discuss VIBE and its application to image databases.

Most current work on video indexing concentrates on queries which operate over high level semantic information which must be entirely composed and entered manually. We propose an indexing system which is based on spatial information about key objects in a scene. These key objects may be detected automatically, with manual supervision, and tracked through a sequence using one of a number of recently developed techniques. This representation is highly compact and allows rapid resolution of queries specified by iconic example. A number of systems have been produced which use 2D string notations to index digital image libraries. Just as 2D strings provide a compact and tractable indexing notation for digital pictures, a sequence of 2D strings might provide an index for a video or image sequence. To improve further upon this we reduce the representation to the 2D string pair representing the initial frame, and a sequence of edits to these strings. This takes advantage of the continuity between frames to further reduce the size of the notation. By representing video sequences using string edits, a notation has been developed which is compact, and allows querying on the spatial relationships of objects to be performed without rebuilding the majority of the scene. Calculating ranks of objects directly from the edit sequence allows matching with minimal calculation, thus greatly reducing search time. This paper presents the edit sequence notation and algorithms for evaluating queries over image sequences. A number of optimizations which represent a considerably saving in search time is demonstrated in the paper.

Data independence is a core property of database management systems (DBMSs) distinguishing them from file-based data management: The application must be presented with information in exactly the form it needs them, without having to transform it in any way before processing and, in particular, independent from the format in which data are stored within the database. For 2D digital images and other raster data such as 1D time series, 3D tomograms, 3D and 4D environmental sensor data and high-dimensional simulation data this means that the application is free to choose between a main memory representation suitable for the target machine type on hand (e.g., to perform a convolution) and some other data format (e.g., to exploit MPEG hardware support). Previously, the concept of Multidimensional Discrete Data (MDD) has been suggested to handle raster data of all kind. A specialized storage architecture has been presented for the generic and efficient storage, manipulation, and retrieval of MDD. In this paper, we use this approach to show how strict separation of logical and physical level together with a declarative query interface leads to full data independence on MDD, as known from the classical DBMS data types such as strings and numbers. At the same time, sufficient flexibility is preserved to support an arbitrary number of specialized formats in parallel. The application can specify that query results shall be delivered as pure, unencoded C/C++ main memory arrays or in any other format implemented in the DBMS which is capable of holding the data. In addition, due to the enhanced semantics available in the database, storage format and database operations can be optimized according to various criteria such as data conversion overhead and transmission bandwidth. Data compression becomes an internal feature invisible to the application and taylorable to each client's actual needs. Benefits are exemplified through an application scenario.

The demand for digital radiological imaging and archiving applications has been increasingly rapidly. These digital applications offer significant advantages to the physician over the traditional film-based technique. They result in faster and better quality services, support remote access and conferencing capabilities, provide on demand service availability, eliminate film processing costs, and most significantly, they are suitable services for the evolving global information super highway. Several existing medical multimedia systems incorporate and utilize those advanced technical features. However, radiologists are seeking an order of magnitude improvement in the overall current system design and performance indices (such as transactions response times, system utilization and throughput). One of the main technical concern radiologists are raising is the miss-filing occurrence. This even will decrease the radiologist productivity; introduce unnecessarily workload; and will result in total customer dissatisfaction. This paper presents Multimedia Medical Archiving System, which can be used in hospitals and medical centers for storing and retrieving radiological images. Furthermore, this paper emphasizes a viable solution for the miss-filing problem. The results obtained demonstrate and quantify the improvement in the overall radiological operations. Specifically this paper demonstrates an order of 80% improvement in the response time for retrieving images. This enhancement in system performance directly translates to a tremendous improvement in the radiologist's productivity.