Search

This paper describes Data Modeling for unstructured data of Diffusion Tensor Imaging (DTI). Data Modeling is an essential first step for data preparation in any data management and data mining procedure. Conventional Entity- Relational (E-R) data modeling is lossy, irreproducible, and time-consuming especially when dealing with unstructured image data associated with complex systems like the human brain. We propose a methodological framework for more objective E-R data modeling with unlimited query support by eliminating the structured content-dependent metadata associated with the unstructured data. The proposed method is applied to DTI data and a minimum system is implemented accordingly. Eventually supported with navigation, data fusion, and feature extraction modules, the proposed system provides a content-based support environment (C-BASE). Such an environment facilitates an unlimited query support with a reproducible and efficient database schema. Switching between different modalities of data, while confining the feature extractors within the object(s) of interest, we supply anatomically specific query results. The price of such a scheme is relatively large storage and in some cases high computational cost. The data modeling and its mathematical framework, behind the scene of query executions and the user interface of the system are presented in this paper.

Protein expression analysis has traditionally relied upon visual evaluation of immunohistochemical reaction by a pathologist, who analyzes the grade of staining intensity and estimates the percentage of cells stained in the area of interest. This method is effective in experienced hands but has potential limitations in its reproducibility due to subjectivity between and within operators. These limitations are particularly pronounced in gray areas where a distinction of weak from moderate protein expression can be clinically significant. Some research also suggests that sub localization of the protein expression into different components such as nuclei versus cytoplasm may be of great importance. This distinction can be particularly difficult to quantify using manual methods. In this paper, we formulate the problem of quantitative protein expression analysis as an active learning classification problem, where a very small set of pre-sampled user data is used for understanding expert evaluation. The expert coveted confidence is mapped to derive an uncertainty region to select the supplemental learning data. This is done by posing a structured query to the unknown data set. The newly identified samples are then augmented to the training set for incremental learning. The strength of our algorithm is measured in its ability to learn with minimum user interaction. Chroma analysis results of a Tissue Micro-array (TMA) images are presented to demonstrate the user interaction and learning ability. The chroma analysis results are then processed to obtain quantitative results.

This paper describes a method for creating agents for locating images of specific categories such as sky, vegetation, fire, and smoke. The method uses only color information, and is based on vector quantization to build a category specific codebook from a set of training images. The method is shown to yield categorization of images collected from several web sites on Internet with a high success rate. The method can be used as an aid to image annotation or as a way to filter images in a content-based image retrieval system.

The goal of the current paper is to introduce a novel clustering algorithm that has been designed for grouping transcribed textual documents obtained out of audio, video segments. Since audio transcripts are normally highly erroneous documents, one of the major challenges at the text processing stage is to reduce the negative impacts of errors gained at the speech recognition stage. Other difficulties come from the nature of conversational speech. In the paper we describe the main difficulties of the spoken documents and suggest an approach restricting their negative effects. In our paper we also present a clustering algorithm that groups transcripts on the base of informative closeness of documents. To carry out such partitioning we give an intuitive definition of informative field of a transcript and use it in our algorithm. To assess informative closeness of the transcripts, we apply Chi-square similarity measure, which is also described in the paper. Our experiments with Chi-square similarity measure showed its robustness and high efficacy. In particular, the performance analysis that have been carried out in regard to Chi-square and three other similarity measures such as Cosine, Dice, and Jaccard showed that Chi-square is more robust to specific features of spoken documents.© (2002) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

In image similarity retrieval systems, color is one of the most widely used features. Users who are not well versed with the image domain characteristics might be more comfortable in working with an Image Retrieval System that allows specification of a query in terms of keywords, thus eliminating the usual intimidation in dealing with very primitive features. In this paper we present two approaches to automatic image annotation, by finding those rules underlying the links between the low-level features and the high-level concepts associated with images. One scheme uses global color image information and classification tree based techniques. Through this supervised learning approach we are able to identify relationships between global color-based image features and some textual decriptors. In the second approach, using low-level image features that capture local color information and through a k-means based clustering mechanism, images are organized in clusters such that images that are similar are located in the same cluster. For each cluster, a set of rules is derived to capture the association between the localized color-based image features and the textual descriptors relevant to the cluster.

In this research, we studied the joint use of visual and audio information for the problem of identifying persons in real video. A person identification system, which is able to identify characters in TV shows by the fusion of audio and visual information, is constructed based on two different fusion strategies. In the first strategy, speaker identification is used to verify the face recognition result. The second strategy consists of using face recognition and tracking to supplement speaker identification results. To evaluate our system's performance, an information database was generated by manually labeling the speaker and the main person's face in every I-frame of a video segment of the TV show 'Seinfeld'. By comparing the output form our system with our information database, we evaluated the performance of each of the analysis channels and their fusion. The results show that while the first fusion strategy is suitable for applications where precision is much more critical than recall. The second fusion strategy, on the other hand, generates the best overall identification performance. It outperforms either of the analysis channels greatly in both precision an recall and is applicable to more general applications, such as, in our case, to identify persons in TV programs.© (2001) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

We proposed an omni-face tracking system for video annotation in this paper, which is designed to find faces from arbitrary views in complex scenes. The face detector first locates potential faces in the input by performing skin-tone detection. The subsequent processing consists of two largely independent components, the frontal face module and the side- view face module, responsible for finding frontal-view and side-view faces, respectively. The frontal face module uses a region-based approach wherein regions of skin-tone pixels are analyzed for gross oval shape and the presence of facial features. In contrast, the side-view face module follows an edge-based approach to look for curves similar to a side-view profile. To extract the trajectories of faces, the temporal continuity between consecutive frames within the video shots is considered to speed up the tracking process. The main contribution of this work is being able to find faces irrespective of their poses, whereas contemporary systems deal with frontal-view faces only. Information regarding to human faces is encoded in XML format for semantic video content representation. The effectiveness of human face for video annotation is demonstrated in a TV program classification system that categories the input video clip into predefined types. It is shown that the classification accuracy is improved saliently by the employment of face information.© (2000) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Currently, most approaches for object tracking are under spatial domain using optical flow and depth, or some model- based methods which need to uncompress those video sequences then do further disposal. The computation for uncompressing is expensive and not good for real time control. Although there are also some researchers doing object tracking under compressed domain, they only use part of DCT values of I frames in a video sequence, which doesn't take advantage of the information under compressed domain. In this paper we consider a new method for object tracking which uses the information only supplied in compressed domain through the MPEG encoder. The main scheme we have performed is to get the motion vectors of P and B frames directly from the MPEG video without decompressing it then cluster objects based on the motion vectors. In particular, camera motion is also taken into account since camera's motion can influence the objects' motion and segmentation results dramatically. Experiments based on the method mentioned above have been carried out in several videos. The results obtained indicate that tracking object in compressed domain is very promising.

In this paper, we do some pre-processing on the input data to remove some noise before putting them into the network and some post-processing before outputting the results. Different neural networks such as back-propagation, radias basis network with different architecture are tested. We choose the one with the best performance among them. From the experiments we can see that the results of the neural network are similar to those given by the experienced doctors and better than those of previous research, indicating that this approach is very practical and beneficial to doctors comparing with some other methods currently existing.

This paper addresses the problem of building a model for text documents of interest. Specifically, it considers a scenario where a large collection of documents, for example theresult of a search on Internet using one of the popular search engines, is given. Each documentis indexed by certain keywords or terms. It is assumed that the user has identified a subset of documents that fits the user's need. The goal is to build a term association model for the documents of interest so that it can be used either for refining the user search or exported toother search engines/agents for further search of documents of interest. The model built is in the form of a unate Boolean function of the terms or keywords used in the initial search of documents. The proposed document model building algorithm is based on a modified version ofthe pocket algorithm for perceptron learning and a mapping method for converting neurons intoequivalent symbolic representations. 1. Introduction The majority of data mining applications deal with situations involving structured data, i.e. datacontaining well-defined alphanumeric fields with unambiguous meanings. This is not surprisingas most business databases contain structured data. In recent years, however, repositories ofsemi-structured or unstructured data, for example databases of text and image documents, haveshown a phenomenal growth, partly due to the popularity of Internet and decreasing hardwarecosts. The text and image data not only lack the structure of the common business data but alsoare ambiguous and generally open to several interpretations of the data. This characteristic isparticularly acute in image and video data. Consequently, management of semi or unstructured

Color is one of the most widely used features for image similarity retrieval. Most of the existing image similarity retrieval schemes employ either global or local color histogramming. In this paper, we explore the use of localized dominant hue and saturation values for color-based image similarity retrieval. This scheme results in a relatively compact representation of color images for similarity retrieval. Experimental results comparing the proposed representation with global and local color histogramming are presented to show the efficacy of the suggested scheme.

Video content characterization is a challenging problem in video databases. The aim of such characterization is to generate indices that can describe a video clip in terms of objects and their actions in the clip. Generally, such indices are extracted by performing image analysis on the video clips. Many such indices can also be generated by analyzing the embedded audio information of video clips. Indices pertaining to context, scene emotion, and actors or characters present in a video clip appear especially suitable for generation via audio analysis techniques of keyword spotting, and speech and speaker recognition. In this paper, we examine the potential of speaker identification techniques for characterizing video clips in terms of actors present in them. We describe a three-stage processing system consisting of a shot boundary detection stage, an audio classification stage, and a speaker identification stage to determine the presence of different actors in isolated shots. Experimental results using the movie A Few Good Men are presented to show the efficacy of speaker identification for labeling video clips in terms of persons present in them.© (1997) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Chroma-keying or blue screen matting is an important video editing operational. In blue screen matting, everything in the image that has the 'user specified level for the blue channel' is 'keyed' out and replaced by either another image or a color from a color generator. We develop a technique for blue screen matting that manipulates the image when the information for the image is available only in compressed form such as a JPEG or MPEG bitstream. Specifically, for the compressed domain approach, we show that the matting process is a convolution operation; hence, we develop a DCT convolution theorem. The DCT convolution theorem can be used to show that the compressed domain approach proposed in this paper provides a significant reduction in computation complexity compared to previously developed approaches which were mostly adhoc techniques. The convolution theorem exploits the sparseness as well as the orthogonality of the data available in the DCT domain and thus yields an efficient algorithm for he chroma-keying or blue screen matting. The algorithm extends to the DCT domain concepts such as alpha channel and premultiplied alpha. The method is also extended to MPEG video domain wherein we explore the efficiency of the matting process when interframe coding is used.© (1997) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

The scanline algorithms are popular in a computer graphics for complex geometric manipulations. The main characteristic of the scanline algorithms is that a geometric transformation is decomposed into multiples transforms with each pass operating only along row or column scanlines. This leads to conversion of 2D image manipulation problems to straightforward 1D problems resulting in simple and systematic methods. The goal of this work is to examine the scanline approach for manipulation of transform-compressed images without decompressing them. We show how the scanline algorithms for rotation and projective mapping can be developed for JPEG/DCT images. The performance of the proposed scanline algorithms is evaluated with respect to quality, speed, and control and memory overhead.

The major problem facing video databases is that of content characterization of video clips once the cut boundaries have been determined. The current efforts in this direction are focussed exclusively on the use of pictorial information, thereby neglecting an important supplementary source of content information, i.e. the embedded audio or sound track. The current research in audio processing can be readily applied to create many different video indices for use in Video On Demand (VOD), educational video indexing, sports video characterization, etc. MPEG is an emerging video and audio compression standard with rapidly increasing popularity in multimedia industry. Compressed bit stream processing has gained good recognition among the researchers. We have also demonstrated feature extraction in MPEG compressed video which implements a majority of scene change detection schemes on compressed video. In this paper, we examine the potential of audio information for content characterization by demonstrating the extraction of widely used features in audio processing directly from compressed data stream and their application to video clip classification.© (1996) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

This paper examines the issue of direct extraction of low level features from compressed images. Specifically, we consider the detection of areas of interest and edges in images compressed using the discrete cosine transform (DCT). For interest areas, we show how a measure based on certain DCT coefficients of a block can provide an indication of underlying activity. For edges, we show using an ideal edge model how the relative values of different DCT coefficients of a block can be used to estimate the strength and orientation of an edge. Our experimental results indicate that coarse edge information from compressed images can be extracted up to 20 times faster than conventional edge detectors.

In recent years, databases have evolved from storing pure textual information to storing multimedia information -- text, audio, video, and images. With such databases comes the need for a richer set of search keys that include keywords, shapes, sounds, examples, sketches, color, texture and motion. In this paper we address the problem of image retrieval where keys are object shapes or user sketches. In our scheme, shape features are extracted from each image as it is stored. The image is first segmented and points of high curvature are extracted. Regions surrounding the points of high curvature are used to compute feature values by comparing the regions with a number of references. The references themselves are picked out from the set of orthonormal wavelet basis vectors. An ordered set of distance measures between each local region and the wavelet references form a feature vector. When a user queries the database through a sketch, the feature vectors for high curvature points on the sketch are determined. An efficient nearest neighbor search then yields a set of images which contain objects that match the user's sketch closely. The process is completely automated. Initial experimental results are presented.

One of the challenging problems in video databases is the organization of video information. Segmenting a video into a number of clips and characterizing each clip has been suggested as one mechanism for organizing video information. This approach requires a suitable method to automatically locate cut points in a video. One way of finding such cut points is to determine the boundaries between consecutive camera shots. In this paper, we address this as a statistical hypothesis testing problem and present three tests to determine cut locations. All the three tests are such that they can be applied directly to the compressed video. This avoids an unnecessary decompression-compression cycle, since it is common to store and transmit digital video in compressed form. As our experimental results indicate, the statistical approach permits accurate detection of scene changes induced through straight as well as optical cuts.

Many applications require similarity retrieval of an image from a large collection of images. In such cases, image indexing becomes important for efficient organization and retrieval of images. This paper addresses this issue in the context of a database of signature images and describes a system for similarity retrieval and recognition of signature images. The proposed system uses a set of geometric and topological features to map a signature image into two strings of finite symbols. A local associative indexing scheme is then used on the strings to organize and search the signature database. The advantage of the local associative indexing is that it is tolerant of missing features and allows queries even with partial signatures. The performance of the system has been tested with promising results with a signature database of 120 signatures.

Many target tracking methods rely on the use of centroids as features points. In several applications such as the tracking of vehicles on roadways, the appropriate target paths are known a priori. Based on this observation, an efficient centroid computation method is presented and analyzed in this paper.

This paper presents a generalization of the correspondence approach of Sethi and Jain by extending the path coherence criterion to high dimensional vector space. This allows the same correspondence procedure to be used for a variety of tokens including points, lines, planes and regions. To demonstrate the generalized approach, we apply it to track lines and present experimental results.

The problem of determining the position and orientation of a mobile robot has been addressed by several researchers using sensors of different modalities, including video cameras. Invariably, all the vision-based approaches for robot localization consider that the camera is mounted on the robot and that the robot working environment is assumed to contain prominent landmarks at known locations. In this paper we propose a robot localization scheme where the robot itself serves as the landmark for cameras that are positioned in the environment to cover the entire work area of the robot. Although the proposed approach is applicable for the robots of any regular shape, we develop the solution to the localization problem by assuming a cylindrical shape for the robot. A compete mathematical analysis of the localization problem is given by extending the three-dimensional structure-from-rotational motion approach to the present task. We also examine the implementation issue of the proposed approach and present experimental results to show its effectiveness.

In addition to invariance with respect to certain geometric transformations, there are two other key requirements for any shape recognition system. It should be flexible enough to adopt to a variety of sets of shapes with minimal training; and it should be capable of performing even in presence of occlusion. This paper describes one such shape recognition system that is currently under development. The system is based on the redundant hashing scheme of Kohonen for recognizing and correcting misspelt words. The current version of the system is meant for 2-D shapes; however, the same approach is applicable to 3-D shape recognition. In the present implementation for 2-D shapes, a polygonal approximation of the given shape is encoded in the form of a string. The encoded string is then used to generate a very small set of shape hypotheses through the use of redundant hashing. The best hypothesis from the set of competing hypotheses is selected by a very simple matching scheme followed by a verification phase based on rotation transformation. The experiments thus far indicate that the system is capable of recognizing shapes in presence of occlusion with about 5% error. The system has significant ability to adopt to new sets of shapes; it does not require any training except the building of hash index table and the corresponding shape dictionary.

Echocardiographic images are generally of poor quality. From a strict gray scale perspective, there are few edges which correspond to the underlying anatomy. Trained ultrasonographers must use considerable a priori information regarding normal human anatomy and use this information extensively to 'fill in' the missing and scanty details of a typical echocardiogram. Due to these constraints, the conventional image processing approaches often fail miserably for echocardiographic images. Knowledge-based approaches for segmenting echo images have been proposed in the recent past; these appear to be much more successful. One of the limitations of knowledge-based approaches is that the acquisition of knowledge and the formation of rules is a relatively difficult task. Moreover, there are instances where experts may have different opinions; this is specially true in the context of echocardiographic images. Additionally, it is much more difficult to codify the expert's knowledge at the image level. Considering these factors plus the typical medical research scenario, where a series of images are acquired at one sitting for a patient, we present in this paper an approach for segmenting echocardiographic image sequences by directly acquiring the expert's knowledge at the image level. The present implementation yields a look up table capturing the expert's knowledge leading to near real time segmentation of echocardiographic images. The results obtained are quite in agreement with the segmentations obtained through trained ultrasonographers.

A critical problem in automatic target recognition is the extraction and tracking of moving targets in high clutter, low resolution, low contrast thermal imagery. This problem is further exacerbated by the large distances that typically exist between the sensor and the targets. In this paper, a new paradigm called motion filtering is presented for exploiting motion to extract and track targets. The main feature of the present approach is the emphasis on the motion information. Experimental results from two sequences of tactical situations are presented to exemplify the capabilities of the proposed approach.

A method for extracting road boundaries using the monochrome image of a visual road scene is presented. The statistical information regarding the intensity levels present in the image along with some geometrical constraints concerning the road are the basics of this approach. Results and advantages of this technique compared to others are discussed. The major advantages of this technique, when compared to others, are its ability to process the image in only one pass, to limit the area searched in the image using only knowledge concerning the road geometry and previous boundary information, and dynamically adjust for inconsistencies in the located boundary information, all of which helps to increase the efficacy of this technique.

In this paper we present a method to segment a range image into regions which correspond to different object surfaces in the scene. We first obtain an equidistance contour map of the range image from slicing the range image at fixed increment distance values. Pixels along a contour are all at about the same distance from the sensor. We have observed that whenever a contour crosses an object surface edge, we would see direction discontinuity, curvature discontinuity, curvature zero-crossing, or termination of the contour. We call these places the critical points of the contour. We divide a contour into segments at its critical points. Next, we find the two corresponding contour segments on two consecutive slices. Every pair of corresponding contour segments defines a small region in the range image. Thus through registering contour segments in consecutive slices we have partitioned a range image into many small regions. Each region corresponds to a portion of an object surface. The last step is to merge these small regions into larger areas based on whether or not the corresponding scene surface segments of two adjacent regions have similar orientations in the 3-D space. The range image segmentation process is completed when the merging process is done. This approach is fast because it analyzes only the pixels along the equidistance contours and the entire process can be completed in just one pass.

extraction. This is in contrast to the early approaches of twoframe analysis. The use of long image sequences permitssome additional constraints like the continuity of motion tobe exploited provided that the frames are acquired at regularshort intervals of time. Motion analysis approaches usinglong image sequences are also more robust to noise due tothe redundancy of input data. This makes these approachesmore suitable for realistic applications.For motion analysis approaches using an extended imagesequence, it is very important that any motion discontinuityacross time be identified. This is independent of the goal ofmotion analysis, as such a detection of motion discontinuitywill permit the integration of desired information only fromthe appropriate part of the extended image sequence. As an

Recently we have proposed a range image segmentation method based on the equidistance contour map extracted from a range image. An equidistance contour of a range image is formed by pixels on the range image whose corresponding scene points are all at a same specified distance from the sensor. We have observed that in different ways the contours reflect the existence of object surface edges, the geometries of object surfaces, and the orientations of these surfaces in the 3-D space. In this paper we present a method which uses the equidistance contours to further characterize the regions obtained from segmentation. This is an integral part of an object recognition process. The meaning of characterization is to find the geometries of the corresponding 3-D object surfaces of the regions from the contours. If a surface has nice analytic form such as planar, spherical, cylindrical, or conical, we determine not only its type but also the values of the parameters which describe its geometry.