Physical (PHY)-layer security has received considerable interest as a way to safeguard data confidentiality and achieve security and privacy in wireless networks. Authentication between two devices is a challenging problem. In this paper, a machine learning algorithm is proposed to detect and identify rogue transmitters relying on a low-dimensional channel feature vector that is obtained by the Karhunen-Loeve transform (KLT). Specifically, a Linde-Buzo-Gray algorithm is designed for improving the reliability and robustness of the proposed scheme, where a Gaussian Mixture Model (GMM) is employed to learn and track the changes of physical layer properties. Simulation results demonstrate that the proposed authentication scheme achieves a higher spoofing detection rate compared to other existing methods.

Video-based activity and behavior analysis of mice has garnered wide attention in biomedical research. Animal facilities hold large numbers of mice housed in "home-cages" densely stored within ventilated racks. Automated analysis of mice activity in their home-cages can provide a new set of sensitive measures for detecting abnormalities and time-resolved deviation from the baseline behavior. Large-scale monitoring in animal facilities requires minimal footprint hardware that integrates seamlessly with the ventilated racks. The compactness of hardware imposes the use of fisheye lenses positioned in close proximity to the cage. In this paper, we propose a systematic approach to accurately estimate the 3D pose of the mouse from single-monocular fisheye-distorted images. Our approach employs a novel adaptation of a structured forest algorithm. We benchmark our algorithm against existing methods. We demonstrate the utility of the pose estimates in predicting mouse behavior in a continuous video.

Video-based activity and behavior analysis for mice has garnered wide attention in biomedical research. Animal facilities hold large numbers of mice housed in `home-cages' densely stored within ventilated racks. Automated analysis of mice activity in their home-cages can provide a new set of sensitive measures for detecting abnormalities and time-resolved deviation from baseline behavior. Large scale monitoring in animal facilities requires minimal footprint hardware that integrates seamlessly with the ventilated racks. Compactness of hardware imposes use of fisheye lenses positioned in close proximity to the cage. In this paper, we estimate the 3D pose of a mouse from fisheye distorted monocular monochromatic images using a novel adaptation of a structured forests algorithm. The method utilizes classification decision trees leveraging their versatility to store arbitrary information in the leaf-nodes. During training, the samples arriving at each node are mapped from continuous pose space to discrete class labels such that similar poses are grouped in the same class. The node splitting function is trained by optimizing a classification objective function rather than a high-dimensional regression one. The leaf-nodes store the pose parameters for the set of samples reaching the node. A prediction model preserving the structural relationship of the pose is formed based on the samples in the leaf-nodes. We apply the method to what we believe is the first known training set for 3D recovery of mouse key points from monocular images. We compare the results of our approach to those obtained via standard regression techniques.

In this paper, we present a novel contrast enhancement method for backlit images that consists of three steps: i) computation of the transmission coefficients using the dark channel prior, ii) generation of multiple images having different exposures based on the transmission coefficients, and iii) image fusion. Compared to global intensity transformation methods and spatially invariant contrast enhancement algorithms, our approach first extracts under-exposed regions using the dark channel prior map, and then performs spatially adaptive contrast enhancement. As a result, the contrast of the image is increased, especially for backlit scenes and those with very wide dynamic range, while still preserving image details and color.

In this paper, we present a dual off-axis color filtered aperture (DCA)-based computational imaging system for depth estimation in a single-camera framework. The DCA has one primary (red) color and its complementary (cyan) color filtered apertures located off the optical axis to generate misalignment between color channels depending on the distance of a region-of-interest from the camera. Disparity of color shifting values (CSVs) between color channels in the image acquired by the DCA-based imaging system is estimated using L1-norm minimization of energy functional considering both brightness and gradient constancies and only translation of x-axis. The two data terms both red to green (R-G) and red to blue (R-B) are combined in a single data term. The proposed imaging system can be implemented by simply inserting an appropriately resized DCA into any general optical system. Experimental results show that the proposed DCA can estimate the distance of the scene from the camera in the single-camera framework.

In this paper, we present an image enhancement algorithm based on the wavelength-adaptive image formation model and geometric classification for defogging UAV images. We first generate a labeled image using geometric class-based segmentation. We then generate a modified transmission map based on the wavelength-adaptive image formation model with scattering coefficients in the labeled image. We also estimate the atmospheric light from the modified transmission map instead of simply choosing the brightest pixel. The proposed method can significantly enhance the visibility of foggy UAV images compared with existing monochrome model-based defogging method. The proposed algorithm can enhance the visibility by removing atmospheric degradation factor in airborne images acquired by aerial platforms such as satellite, airplane, and UAV under critical weather conditions such as haze, fog, and smoke.

A multiple color-filter aperture (MCA) camera can provide depth information as well as color and intensity in the single-camera framework, where the MCA generates misalignment between color channels depending on the distance of a region-of-interest. In this paper, we present a single camera-based estimation of the full depth map using the color shifting property of the MCA. For estimating the color shifting vectors (CSVs) among red, green, and blue color channels, edges are extracted at each color channel. At the edge, we estimate CSVs using normalized cross correlation combined with color shifting mask map. A full depth map is then generated by depth interpolation using the matting Laplacian method from sparsely estimated CSVs at an edge location. Experimental results show that the proposed method can not only estimate the full depth map but also correct the misaligned color image to generate photorealistic color images using a single camera equipped with MCA.

This paper presents a spatially adaptive super-resolution (SR) algorithm using homogeneous region analysis for minimizing undesired interpolation artifacts such as aliasing and jagged edges. The proposed regularized SR algorithm incorporates two constraints enforcing (i) disconnected edges to be reconnected and (ii) orientation-adaptive regularization. By combining two constraints in the regularization framework, the proposed SR algorithm can significantly reduce aliasing artifacts and, as a result, produce edge-preserved high-resolution (HR) images. In addition to the formulation of the regularized SR algorithm with hybrid constraints, experimental results show that the proposed SR algorithm improves peak-to-peak signal-to-noise ratio (PSNR) of the reconstructed HR images by up to 5[dB]over existing state-of-the-art SR methods.

This paper presents a novel approach to depth estimation using a multiple color-filter aperture (MCA) camera and its application to multifocusing. An image acquired by the MCA camera contains spatially varying misalignment among RGB color channels, where the direction and length of the misalignment is a function of the distance of an object from the plane of focus. Therefore, if the misalignment is estimated from the MCA output image, multifocusing and depth estimation become possible using a set of image processing algorithms. We first segment the image into multiple clusters having approximately uniform misalignment using a color-based region classification method, and then find a rectangular region that encloses each cluster. For each of the rectangular regions in the RGB color channels, color shifting vectors are estimated using a phase correlation method. After the set of three clusters are aligned in the opposite direction of the estimated color shifting vectors, the aligned clusters are fused to produce an approximately in-focus image. Because of the finite size of the color-filter apertures, the fused image still contains a certain amount of spatially varying out-of-focus blur, which is removed by using a truncated constrained least-squares filter followed by a spatially adaptive artifacts removing filter. Experimental results show that the MCA-based multifocusing method significantly enhances the visual quality of an image containing multiple objects of different distances, and can be fully or partially incorporated into multifocusing or extended depth of field systems. The MCA camera also realizes single camera-based depth estimation, where the displacement between multiple apertures plays a role of the baseline of a stereo vision system. Experimental results show that the estimated depth is accurate enough to perform a variety of vision-based tasks, such as image understanding, description, and robot vision.

A new night-time lane detection system and its accompanying framework are presented in this paper. The accompanying framework consists of an automated ground truth process and systematic storage of captured videos that will be used for training and testing. The proposed Advanced Lane Detector 2.0 (ALD 2.0) is an improvement over the ALD 1.0 or Layered Approach with integration of pixel remapping, outlier removal, and prediction with tracking. Additionally, a novel procedure to generate the ground truth data for lane marker locations is also proposed. The procedure consists of an original process called time slicing, which provides the user with unique visualization of the captured video and enables quick generation of ground truth information. Finally, the setup and implementation of a database hosting lane detection videos and standardized data sets for testing are also described. The ALD 2.0 is evaluated by means of the user-created annotations accompanying the videos. Finally, the planned improvements and remaining work are addressed.

In this paper, we propose a new methodology for detecting lane markers that exploits the parallel nature of lane boundaries on the road. First, the input image is pre-processed and Altered to detect lane marker features. Then, using the Polar Randomized Hough Transform that is introduced in this paper, lines are fitted through the detected features and the orientation of each line is evaluated. By finding near parallel lines separated by a constraint specified distance, false signalling caused by artifacts in the image is greatly reduced. The proposed system was tested using a real world driving videos and showed good results despite the presence of neighboring vehicles, shadows, and irregularities on the road surface.

Recognizing handwritten mathematical content is a challenging problem, and more so when such content appears in classroom videos. However, given the fact that in such videos the handwritten text and the accompanying audio refer to the same content, a combination of a video and an audio based recognizer has the potential to significantly improve the content recognition accuracy. In this paper, using a combination of video and audio based recognizers, we focus on improving the character recognition accuracy in such videos and propose: (1) synchronization techniques for establishing a correspondence between the handwritten and the spoken content, and (2) combination techniques for combining the outputs of the video and audio based recognizers. The current implementation of the system makes use of a modified open source text recognizer and a commercially available phonetic word-spotter. For evaluation purposes, we use videos recorded in a classroom-like environment and our experiments demonstrate the significant improvements (≈ 24% relative increase as compared to the baseline video based recognizer) in character recognition accuracy that can be achieved using our techniques.

In this document, a new and efficient method to specify the ground truth locations of lane markers is presented. The method comprises of a novel process called Time-Slicing that provided the user with a unique visualization of the video. Coupled with automation via spline interpolation, the quick generation of necessary ground truth information is achieved. Videos recorded from a vehicle while driving on local city roads and highways are marked with ground truth information for use in testing. The performance of a variety of lane detection systems is compared to the ground truth and the error is computed for each system. Finally, quantitative analysis shows that the reference lane detection system presented in produces the most accurate lane detections which is depicted by the smallest error.

In this paper, we present a new methodology for detecting lane markers that is able to withstand many challenging situations like scattered shadows, illumination changes, and presence of neighboring vehicles to name a few. At first, the input image undergoes a perspective removal followed by a color space conversion. Then, the core elements consisting of template matching, lane region merging, and elliptical projections are explained. Finally, the developed system showed good results when tested on 15 minutes of real-world driving videos containing variations in illumination, traffic, and road surface conditions.

In a previous paper, a simple approach to lane detection using the Hough transform and iterated matched filters was described. This paper extends this work by incorporating an inverse perspective mapping to create a bird's-eye view of the road, applying random sample consensus to help eliminate outliers due to noise and artifacts in the road, and a Kalman filter to help smooth the output of the lane tracker.

While the ongoing information technology (IT) revolution is providing us with tremendous educational opportunities, educators and IT researchers face numerous obstacles and pedagogical questions. The Georgia Institute of Technology (Georgia Tech) has a long history in engineering education research and has developed and designed many different tools for instructor authoring, course capturing, indexing, and retrieval. Special attention has been applied to the design and deployment of distributed learning environments. This paper describes the environment and challenges facing Georgia Tech as it expands its campus worldwide while maintaining integration of faculty and students across these campuses. The focus is primarily on the problems of synchronous delivery to multiple sites with a description of how technology is currently being deployed in Georgia Tech's distance-learning (DL) classrooms, as well as technology that is under development for the DL classroom of the future.

The paper describes an application of practical technologies to implement a low cost, consumer grade, single chip biometric system based on face recognition using infra-red imaging. The paper presents a system that consists of three stages that contribute in the face detection and recognition process. Each stage is explained with its individual contribution alongside results of tests performed for that stage. The system shows a high recognition rate when full frontal face images are led to the system. The paper further discusses the application based approach in the automotive world with plans for further study. Recognition rates of the overall system are also presented.

The active appearance model (AAM) is a powerful tool for modeling images of deformable objects and has been successfully used in a variety of alignment, tracking, and recognition applications. AAM uses subspace-based deformable models to represent the images of a certain object class. In general, fitting such complicated models to previously unseen images using standard optimization techniques is a computationally complex task because the gradient matrix has to be numerically computed at every iteration. The critical feature of AAM is a fast convergence scheme which assumes that the gradient matrix is fixed around the optimal coefficients for all images. Our work in this paper starts with the observation that such a fixed gradient matrix inevitably specializes to a certain region in the texture space, and the fixed gradient matrix is not a good estimate of the actual gradient as the target texture moves away from this region. Hence, we propose an adaptive AAM algorithm that linearly adapts the gradient matrix according to the composition of the target image's texture to obtain a better estimate for the actual gradient. We show that the adaptive AAM significantly outperforms the basic AAM, especially in images that are particularly challenging for the basic algorithm. In terms of speed and accuracy, the idea of a linearly adaptive gradient matrix presented in this paper provides an interesting compromise between a standard optimization technique that recomputes the gradient at every iteration and the fixed gradient matrix approach of the basic AAM.

This paper proposes a novel system that can automatically create an optimal and nonrepetitive summarization and support different user requirements for video browsing and content overview by outputting both the optimal set of key frames and a summarized version of the original video with the user-specified time length. Comparing our approach to video abstraction with another algorithm, we demonstrate that our approach is fast and produces an effective video summary.

The objective of this work is to provide a simple and efficient method to detect human faces in the compressed domain. We separate skin regions from non-skin regions using the Bayesian decision rule with a Gaussian mixture model (GMM). Next, we detect the location of a face using a deformable template that considers shape and orientation of the face. Then, face candidates are verified by using the second moments computed directly from the DCT coefficients in the face region. In particular, the coefficients are manipulated differently according to the orientation of the face. Good results have been obtained for a large variety of video sequences. This algorithm can be applied to JPEG images as well, without any modification.

We introduce some interesting properties of the singular value decomposition (SVD), and illustrate how they may be used in conjunction with the k-means algorithm for efficiently clustering a set of vectors. Specifically, we use the SVD to preprocess and sort the data vectors, and then use the k-means algorithm on the modified vectors. To illustrate the effectiveness of this approach, we compare it to the k-means algorithm without preprocessing and show that significant gains in clustering speed may be realized.

The active appearance model (AAM) algorithm is a powerful tool for modeling images of deformable objects. AAM combines a subspace-based deformable model of an object's appearance with a fast and robust method of fitting this model to a previously unseen image. The speed of this algorithm comes from the assumption that the gradient matrix is fixed around the optimal coefficients for all images. In this paper, we propose a convergence scheme for AAM that adapts this gradient matrix to the target image's texture during convergence by adding linear modes of change that are based on the texture eigenvectors of AAM. We show that this adaptive strategy for the gradient matrix provides a significant increase in the performance of the AAM algorithm.

Face recognition is an important part of today's emerging biometrics and video surveillance markets. Recent years have witnessed an exploding interest in the development of face recognition algorithms and products. Currently, face recognition systems are usually implemented on general purpose processors. As face recognition algorithms move from research labs to the real world, power consumption and cost become critical issues. This motivates searching for implementations using a digital signal processor (DSP). Our goal is to explore the feasibility of implementing DSP-based face recognition systems. To achieve this goal, we implement a fully automatic face recognition system on Texas Instruments' TMS320C6416 DSP, profile performance, and analyze opportunities for optimization. The results of our experiments demonstrate that a generic C implementation with a modest C level optimization effort results in a face recognition software prototype that has low CPU and memory requirements. Hence, it appears that well-optimized face recognition implementations on DSPs can be an effective choice for embedded face recognition products.

This paper introduces a useful property of the singular value decomposition (SVD) and uses it to quickly summarize a video sequence based on the visual similarities of its frames. In our method, a video is expressed as the representative frames extracted by a simple key-frame extraction algorithm applied in a sequential manner. Then those key-frames are put together with little redundancy using a clustering algorithm for video abstraction. In order to evaluate the proposed scheme, the speed of the commonly used k-means algorithm for clustering is compared with that of the proposed method that combines both the SVD and the k-means algorithm. Experimental results show that our algorithm is fast and effectively summarizes the content of a video with little redundancy.

Face images that are captured by surveillance cameras usually have a very low resolution, which significantly limits the performance of face recognition systems. In the past, super-resolution techniques have been proposed to increase the resolution by combining information from multiple images. These techniques use super-resolution as a preprocessing step to obtain a high-resolution image that is later passed to a face recognition system. Considering that most state-of-the-art face recognition systems use an initial dimensionality reduction method, we propose to transfer the super-resolution reconstruction from pixel domain to a lower dimensional face space. Such an approach has the advantage of a significant decrease in the computational complexity of the super-resolution reconstruction. The reconstruction algorithm no longer tries to obtain a visually improved high-quality image, but instead constructs the information required by the recognition system directly in the low dimensional domain without any unnecessary overhead. In addition, we show that face-space super-resolution is more robust to registration errors and noise than pixel-domain super-resolution because of the addition of model-based constraints.

In this paper, a simple approach to camera motion analysis is proposed to index videos compressed using MPEG (Moving Picture Experts Group) −1, 2 in faster than real time. Specifically, this paper presents a template-matching algorithm to classify basic camera operations. The proposed approach involves 1) the construction of motion vector fields (MVFs) and filtering out noise; and 2) the sub-MVFs which come from dividing an MVF into several non-overlapping areas for the template matching of camera motions. A fine segmentation also can be obtained for a video, based on the homogeneity of the camera operation in each unit. The advantages of this method lie in real-time processing and robustness to noisy environments such as false motion vectors and object motions. The experimental results show that the proposed algorithm successfully classifies the camera movements.

Face images that are captured by surveillance cameras usually have a very low resolution, which significantly limits the performance of face recognition systems. In the past, super-resolution techniques have been proposed that attempt to increase the resolution by combining information from multiple images. These techniques use super-resolution as a preprocessing system to obtain a high resolution image that can later be passed to a face recognition system. Considering that most state-of-the-art face recognition systems use an initial dimensionality reduction method, we propose embedding the super-resolution algorithm into the face recognition system so that super-resolution is not performed in the pixel domain, but is instead performed in a reduced dimensional domain. The advantage of such an approach is a significant decrease in the computational complexity of the super-resolution algorithm because the algorithm no longer tries to construct a visually improved high quality image, but instead constructs the information required by the recognition algorithm directly in the lower dimensional domain without any unnecessary overhead.

A fast means of object segmentation in a video sequence using low-level features existing in the compressed video stream is presented. These features include the DCT coefficient values of I-frames and motion vectors. The work described here is the foundation of a spatial segmentation system that approaches real time. Potential applications for the system include the separation of foreground and background objects and video database searching.

In this paper, we present a segmented linear subspace model for face recognition that is robust under varying illumination conditions. The algorithm generalizes the 3D illumination subspace model by segmenting the image into regions that have surface normals whose directions are close to each other. This segmentation is performed using a K-means clustering algorithm and requires only a few training images under different illuminations. When the linear subspace model is applied to the segmented image, recognition is robust to attached and cast shadows, and the recognition rate is equal to that of computationally more complex systems that require constructing the 3D surface of the face.

The recent rapid increase in applications of DSP has generated a need for technical personnel with a DSP background. Until recently, most DSP education has been at the graduate level, so many practicing engineers are developing a need for additional DSP education. At Georgia Institute of Technology, we are delivering an online continuing education course that is targeted at these engineers. The 12-week course is comprised of three interconnected tracks in DSP system theory, real-time implementation principles, and laboratory assignments with MATLAB and the latest fixed-point TI DSP chips. The course was designed after a needs analysis performed by Texas Instruments training specialists. It is intended specifically to provide a full semester's course in a way that is adapted to the schedules and circumstances of practicing engineers. We describe both the content of the course and the technology for delivering it online.

Content-based indexing and retrieval of video have a great role in many applications, and scene change detection plays an essential role in these tasks. In this paper, we introduce an algorithm for sudden and gradual scene change detection in MPEG video in faster than real-time. First, we use the number of bi-directional predicted macroblocks (bi-MBs) within a B frame and their mean to define an adaptive threshold to identify abrupt scene changes. Then, a selected frame consisting of MBs is divided into several sub-MB images to detect gradual scene changes, such as dissolves and fades. This algorithm can be applied to any type of MPEG sequence. Experimental results show that this algorithm is fast and robust to detect most abrupt and gradual scene changes in compressed sequences. This method can be used as pre-processing for indexing and retrieval.

Content-based indexing and retrieval of video are important in many applications, and scene-change detection plays an essential role in these tasks. We introduce a faster than real-time algorithm for sudden and gradual scene change detection in MPEG video. For sudden scene change detection, we use an adaptive threshold that is applied to the number of bi-directional predicted macroblocks within a B-frame. To direct gradual scene changes such as dissolves and fades, we divide a selected frame into sub-macroblock images, and perform an analysis of the motion vectors within this image. To refine the gradual scene change, a camera motion analysis technique is performed. Experimental results show that this algorithm is fast and robust for the detection of most abrupt and gradual scene changes in the compressed domain.

The embedded hidden Markov model (HMM) is a statistical model that can be used in many pattern recognition and computer vision applications. This model inherits the partial size invariance of the standard HMM, and, due to its pseudo two-dimensional structure, is able to model two-dimensional data such as images, better than the standard HMM. We describe the maximum likelihood training for the continuous mixture embedded HMM and present the performance of this model for face detection and recognition. The experimental results are compared with other approaches to face detection and recognition.

Stereo image pair coding is an important issue in stereo data compression. A wavelet based stereo image pair coding algorithm is proposed in this paper. The wavelet transform is used to decompose the image into an approximation image and three edge images. In the wavelet domain, a disparity estimation technique is developed to estimate the disparity field using both approximation image and edge images. To improve the accuracy of estimation of wavelet images produced by the disparity compensation technique, a novel wavelet based subspace projection technique (SPT) is developed. In the SPT, the block dependent subspaces are constructed using block varying basis vectors that are derived from the disparity compensated wavelet images. Experimental results show that the proposed algorithm is efficient to achieve stereo image compression.

We describe an embedded hidden Markov model (HMM)-based approach for face detection and recognition that uses an efficient set of observation vectors obtained from the 2D-DCT coefficients. The embedded HMM can model the two dimensional data better than the one-dimensional HMM and is computationally less complex than the two-dimensional HMM. This model is appropriate for face images since it exploits an important facial characteristic: frontal faces preserve the same structure of "super states" from top to bottom, and also the same left-to-right structure of "states" inside each of these "super states".

Our goal is to exploit human motion and object context to perform action recognition and object classification. Towards this end, we introduce a framework for recognizing actions and objects by measuring image-, object- and action-based information from video. Hidden Markov models are combined with object context to classify hand actions, which are aggregated by a Bayesian classifier to summarize activities. We also use Bayesian methods to differentiate the class of unknown objects by evaluating detected actions along with low-level, extracted object features. Our approach is appropriate for locating and classifying objects under a variety of conditions including full occlusion. We show experiments where both familiar and previously unseen objects are recognized using action and context information.

In this paper, we present the notion of cascading steerable filters to improve their angular resolution. Additionally, we illustrate that the results of such cascades can be steered themselves. An advantage of this approach is that only a single, relatively small set of steerable filters can be employed to achieve various angular resolutions. Improving angular resolution has previously required an entirely different, larger set of filters.

The Georgia Institute of Technology in Atlanta is in a unique position of having a campus located in the heart of the Lorraine region of France in the city of Metz. Since the number of faculty at GTL is limited, some courses must be offered in non-traditional formats. We describe some of our current approaches to Internet course delivery, which involve the use of streaming media. Although there are many advantages to Internet distance learning, we describe what we see are the most important problems and pitfalls and present our view of the future of Internet distance learning.

The work presented in this paper focuses on the use of hidden Markov models for face recognition. A new method based on the extraction of 2D-DCT feature vectors is described, and the recognition results are compared with other face recognition approaches. The method introduced reduces significantly the computational complexity of previous HMM-based face recognition system, while preserving the same recognition rate.

The work presented in this paper describes a hidden Markov model (HMM)-based framework for face recognition and face detection. The observation vectors used to characterize the states of the HMM are obtained using the coefficients of the Karhunen-Loeve transform (KLT). The face recognition method presented reduces significantly the computational complexity of previous HMM-based face recognition systems, while slightly improving the recognition rate. Consistent with the HMM model of the face, this paper introduces a novel HMM-based face detection approach using the same feature extraction techniques used for face recognition.

The Georgia Institute of Technology in Atlanta is in a unique position of having a campus located across the Atlantic in the heart of the Lorraine region of France in the city of Metz. With Masters Degree programs in electrical and mechanical engineering, Georgia Tech Lorraine (GTL) offers more courses than can be staffed by resident faculty. Therefore, GTL offers a number of courses in electrical engineering (EE), mechanical engineering (ME), and mathematics by videotaping courses that are taught by faculty at the Georgia Tech campus in Atlanta. Delivering courses by videotape, however, is relatively expensive, introduces a two week delay between the two sections of the courses, and is not the most convenient method of course delivery for the student. In this paper, we describe some approaches that we have used in delivering graduate level EE courses in near real-time across six time zones using the Internet.

An alternative form of the fractal interpolation function (FIF)-previously unmentioned in the signal processing literature-is noted. This form highlights a simple relationship between fractal and linear interpolation. Using this relationship, many FIF problems can be reduced to a matrix/vector expression. This expression provides a more powerful way to employ the FIF for interpolation and permits its adaptation for reconstruction. Additionally, the alternate form of the FIF allows the construction of fractal functions whose piecewise integrals match observed data.

Recently, due to advances in display technology, three-dimensional (3-D) imaging systems are becoming increasingly popular. One way of stimulating 3-D perception is to use stereo pairs, a pair of images of the same scene acquired from different perspectives. Since there is an inherent redundancy between the images of a stereo pair, data compression algorithms should be employed to represent stereo pairs efficiently. This paper focuses on the stereo image coding problem. We begin with a description of the problem and a survey of current stereo coding techniques. A new stereo image coding algorithm that is based on disparity compensation and subspace projection is described. This algorithm, the subspace projection technique (SPT), is a transform domain approach with a space-varying transformation matrix and may be interpreted as a spatial-transform domain representation of the stereo data. The advantage of the proposed approach is that it can locally adapt to the changes in the cross-correlation characteristics of the stereo pairs. Several design issues and implementations of the algorithm are discussed. Finally, we present empirical results suggesting that the SPT approach outperforms current stereo coding techniques.

We discuss several techniques for stereo image coding. These techniques include those that are directly adapted from motion compensation such as disparity compensation, windowed disparity compensation, and variable block size disparity compensation (quadtree decomposition) as well as a novel stereo coding algorithm that we have previously proposed, namely the subspace projection technique (SPT). Also, combined techniques such as windowed disparity compensation followed by the SPT or variable block size disparity compensation followed by the SPT are implemented. In addition, improvements and modifications for the subspace projection technique are reported. These improvements include a non-statistical significance test for the subspace coefficients, two different implementations of the basic algorithm, a locally optimal bit-allocation scheme, and a new quantization scheme for the projection coefficients. In addition, we compare the rate-distortion performance of our approach to several stereo coding algorithms.

We propose a new framework for the compression of multi-view image sequences. We define three types of frames, and each type is coded with a different strategy. The first type of frame is independently coded and is called I-frame. The second is a B-frame and is coded using a bidirectional disparity estimator and a modified version of the subspace projection technique (SPT). The SPT algorithm compensates the photometric variations between the multi-view frames. The projection block size is chosen to be small so that coding of the residual image is not necessary. On the other hand, to decrease the overhead information both disparity vectors and projection coefficients are coded with a lossy scheme. Finally, the third type of frame is a P-frame and is coded by employing a unidirectional disparity estimator and DC level compensation.

An approach is presented for the compression of motion video sequences using iterated function systems. In the proposed approach, the video stream is partitioned into three-dimensional range regions. Each range region consists of a variable number of rectangular blocks that belong to consecutive frames along motion trajectory. Our approach exploits correlation between consecutive blocks in the direction of motion by predicting the IFS map of a given range block with that of a parent range block along the trajectory of motion. The proposed approach shows good promise for efficient modeling and compression of motion video sequences at an affordable computational cost.

Human facial expression modeling has been an active research area. Most of the existing systems do not provide an easy way to adjust the model such that different levels of detail of expressions can be modeled. We propose a method for modeling human facial expressions at different resolutions. Our method is based on the facial action coding system (FACS) developed by Ekman and Freisen (1977). Although the facial expressions we are mainly interested in are those related to speech either directly or indirectly, the modeling method can be extended to apply to facial expressions that are not related to speech.

A region-based stereo image coding algorithm is proposed and evaluated. Three types of regions are considered: occlusion, edge and smooth regions. The region in the right image that is occluded due to finite viewing area is estimated using a new approach. This region is independently coded. The non-occluded region is segmented into edge and smooth regions. Each region is composed of fixed size blocks. The disparity for each block in a non-occluded region is estimated using a block-based approach. The estimated disparity field is encoded employing a lossy residual uniform scalar quantizer and an adaptive arithmetic coder based on the segmentation information. The decoded vectors are used for the subspace projection technique, which is a combined disparity and illumination compensation algorithm. The proposed approach is shown to outperform standard block-based and independent coding algorithms.

An approach to stereo image coding based on disparity compensation and subspace projection is proposed and evaluated. Traditional stereo image coding techniques employing block-based disparity compensation have problems with occlusion regions and photometric variations. We show that the results of these algorithms can be improved if the subspace projection technique (SPT) is employed as a post processing technique after disparity estimation. SPT is an orthogonal basis approach. It combines block-based disparity estimation with two-dimensional first order approximation. For the regions where disparity estimation fails, such as occlusion regions, first order approximation provides a decent estimate. On the other hand, zeroth order term compensates for photometric variations. Empirical results verified that SPT performs better than the disparity compensation in the rate-distortion sense.

Three dimensional display of video sequences with look-around capability enhances realism and introduces a unique sense of "being there". This final step of display realism can be achieved by stereoscopic techniques with the use of multi-view image sequences. A multi-view image system is an extended stereoscopic system with look-around capability. One technical problem that keeps these systems from widespread application is the large amount of information contained in a multi-view system. This information is generally 5 to 10 times more than the information content of an image. However, there is high inter-frame correlation in a multi-view image set which can be exploited for compression purposes. Therefore, the authors investigate efficient storage and transmission techniques for multi-view images. A block-based disparity compensated coder is proposed. This proposed algorithm solves the occlusion problem and is a good candidate for real-time implementation.>

An approach is presented for the compression of motion video sequences using iterated function systems. In the proposed approach, the video stream is partitioned into three-dimensional range regions. Each range region consists of a variable number of rectangular blocks that belong to consecutive frames along motion trajectory. The approach exploits correlation between consecutive blocks in the direction of motion by predicting the IFS map of a given range block with that of a parent range block along the trajectory of motion. The proposed approach shows good promise for efficient modeling and compression of motion video sequences at an affordable computational cost.>

Iterated Function Systems (IFS's) are receiving a lot of attention in the literature as a new technique for signal and image coding. Much of this interest stems from the fact that an IFS is simple in form and yet capable of representing complicated signals, many of which closely resemble those found in nature. In this paper, we introduce the basic ideas involved in using iterated function systems for image coding and present a new approach that is based on the construction of an orthonormal basis to represent the image blocks and then uses simple projections to code each block.>

A pel-recursive two-dimensional motion estimation algorithm using three or more image frames is presented. It uses more information from the image sequences and gives a closed-form solution to update motion vectors among frame pairs simultaneously. Compared to other multiframe-based algorithms, the proposed algorithm is in a more general form. It is suitable for object motion with acceleration since it places no constraints on the displacement vectors among consecutive frames.>

The authors consider the phase retrieval problem using a window function w(n). A condition is developed under which a desired signal can be defined to within a sign ambiguity from the Fourier intensities of the desired signal, the desired signal windowed by w(n), and the desired signal windowed by the window w/sub c/(n)=1-w(n). An algorithm is developed that can reconstruct the desired signal from the given information. An example that shows the performance of the algorithm and two examples that show that, for a certain type of window, two Fourier intensities may be enough for the reconstruction of a signal are presented.>

A generalized symbolic signal representation capturing the physical appearance of process monitoring signals is introduced. The signals are decomposed into simpler subcomponents described by different parametric models and combined under an object-oriented hybrid signal representation. A set of subcomponents is modeled by a piecewise spline function approximation. This model represents the signal patterns in purely mathematical terms and can be viewed as an extension of syntactic signal modeling approaches used in electrocardiogram (ECG) analysis. Furthermore, a robust event-synchronous system capable of transcribing given signals into the proposed symbolic representation is described. The use of event-synchronous analysis makes the system specifically suited for process monitoring applications. A concrete example of a pump monitoring application is described.>

A motion compensated video coding system that minimizes the overall system output bit rate by jointly compressing motion compensated difference (MCD) and motion displacement field (MDF) images is proposed. The MDF constructed from the set of motion vectors is treated as a vector-pixel image and coded to reduce the number of displacement vectors. The bit rates needed to adequately represent MCD and MDF images are interdependent. Thus, for optimal performance a motion compensated coding systems, both MCD and MDF images should be treated as a single unit. Based on the characteristics of these two types of images, a segmentation-based coder is developed. A split-and-merge segmentation algorithm is developed in a graph theoretic context to minimize a coding distortion function. Simulation results have demonstrated excellent performance from this coding system.>

The authors introduce several techniques which are used in the X-ray crystallography phase retrieval problem and several unique problems associated with it. Phase retrieval involves signal reconstruction from its Fourier transform magnitude, which in this case is related to the X-ray diffraction pattern intensity. The following techniques are discussed: the isomorphic addition technique, the isomorphic replacement technique, and the multiple anomalous dispersion technique. Finally, a simulation result for the X-ray crystallography problem is presented.>

A model which may be used to determine the iterated functions system parameters needed to produce a fractal in R/sup 3/ where the projection of that fractal to R/sup 2/ matches a given function is presented. Background material is presented in which the form of the model, model parameters, and fractal synthesis are given. An inverse algorithm for identification of the model parameters is presented, and comments on this algorithm are given. Applications of hidden-variable fractal interpolation to two data types, monthly relative number of sunspots and the electrocardiogram of a dog's heart, are presented.>

Two-dimensional arrival angles of narrowband signals impinging on a volume array are estimated. Direction-of-arrival (DOA) angles are computed by using a master volume array which can be treated as being composed of overlapping subarrays. Considering the entire array in this way, it is possible to estimate the directions of arrival for up to (M-4) sources (e.g. with a triangular prism array configuration) where M is the number of sensors. By using the orthogonality property of the signal and noise subspaces and the definition of the DOA matrix, it is proven that the employment of overlapping subarrays increases the detectability without increasing computation. The 2-D DOA angles may be determined using low observation time as well as low signal-to-noise ratios.>

The problem of reconstructing either a one-dimensional or a two-dimensional signal from its Fourier intensity and the intensity of another signal that contains a known reference signal is addressed. After a brief summary of some conditions under which a one-dimensional or two-dimensional signal can be uniquely defined in terms of the given Fourier intensities, an iterative algorithm that makes it possible to reconstruct the original signal is presented. For the case of a point reference signal, a method is presented that can be used to determine the location and amplitude of the reference from the given Fourier intensity information. An example that illustrates the performance of the reconstruction algorithm is given.>

Levinson's recursion allows for the parameterization of an all-zero or an all-pole filter in terms of a sequence of reflection coefficients. Such a parameterization leads to a simple stability test for filters and serves as the foundation for the lattice filter structure. The properties of the mapping from the reflection coefficient sequence to the system function are considered. It is shown that, in many cases, they parallel the properties of the mapping from the filter coefficients to the system function.>

For a signal consisting of p complex exponentials in white noise, it is well known that the true covariance matrix will have the Hermitian Toeplitz structure and that its minimum eigenvalue will have a dimension of M-p (where M is the dimension of the matrix). When the covariance matrix is estimated from such a signal, it will not generally satisfy these constraints. Two algorithms are presented for imposing these constraints on a covariance matrix. It is shown that the second algorithm generalizes easily to the two-dimensional case. Examples are given to demonstrate the improvement that these algorithms offer for the harmonic retrieval problem. >

At the ICASSP'86 conference in Tokyo we introduced a new efficient ARMA lattice structure that is fully consistent with the geometrical characteristics of the AR and MA lattice structures in that it is realized in terms of a fully orthogonal set of lattice vectors and it evaluates all optimal ARMA filters of lower order than a given maximum order. In that same paper we briefly described a class of RLS algorithms for the evaluation of the ARMA lattice filter coefficients. In this paper that class of algorithms is expanded and the numerical properties of the algorithms are investigated. Practical implementation issues are also discussed.

A generalized iterative deconvolution algorithm is presented. It is shown that the algorithm can be tailored to achieve a pth-order rate of convergence to the inverse filter solution for any integer p. When compared to the method of successive approximations, the generalized algorithm is computationally more efficient in obtaining the same solution.

In this paper a novel data adaptive power spectrum estimation procedure is introduced. This estimate, called PSEUDA for Power Spectrum Estimation Using Data Adaptation, unifies under one conceptual framework the Autoregressive or Maximum Entropy Method (MEM) and the Maximum Likelihood Method (MLM) by providing a method which is continuosly variable between these two extremes For small bandwidths the results obtained are similar to the Data Adaptive Spectral Estimate (DASE) introduced by Davis and Regier [4]. Unlike DASE, a fast algorithm has been developed for the PSEUDA estimate. This algorithm consist of three parts: 1) a Levinsion recursion, 2) a weighted correlation of the prediction error coefficients obtained in step one, and 3) the use of an FFT to calculate the estimate of the spectrum.

This paper looks at least squares ARMA modeling of linear time varying systems with lattice filters. The modeling problem is formulated in a Hilbert space as it is an enlightening approach, that provides very powerful orthogonality relations to work with. There are two parts to this paper. The first part presents a new ARMA lattice filter ARMA(N,M) which is fully consistent with the geometrical characteristics of the AR and MA lattice filters in that it is realized in terms of a fully orthogonal lattice basis and it evaluates all optimal ARMA(i,j) filters of lower order. It therefore goes further in the basis orthogonalization than the ARMA lattice of Lee, Friedlander and Morf and it does not require that N=M. The second part of the paper presents a new fast RLS algorithm for the evaluation of the lattice filter coefficients. The algorithm is based on an inner product factorization and differs from other RLS lattice algorithms in that the projection of the so called pinning vector does not appear in any of the time updates. The algorithm is formulated as a sliding window algorithm, but it embeds a growing window (prewindowed) algorithm which is realized simply by dropping terms from the sliding window algorithm.

In this paper, the application of a constrained iterative deconvolution algorithm to the problem of enhancing noncoherent radar data is described. It is shown that noncoherent reconstruction is theoretically not unique. Despite this, an approximate convolutional model of the observed data is established and it is seen that reasonable results are observed. A key feature of the algorithm is the use of a median filter and threshold to segment the data into two classes, potential point targets and background, which are deconvolved separately subject to different nonlinear constraints. The performance of the algorithm is demonstrated using real air-to-ground millimeter-wave noncoherent radar data.

The harmonic decomposition of a random process into a sum of sinusoids in white noise is an important problem with applications in a number of different areas. As a result of the work of V.F. Pisarenko, it has been shown that the sinusoidal frequencies and the white noise power are determined by the minimum eigenvalue and the corresponding eigenvector of the autocorrelation matrix. Many authors have formulated adaptive versions of the Pisarenko solution for the purpose of spectral line tracking, but most have concentrated only on stationary sinusoids. In most of the interesting applications, however, the frequencies of the sinusoids change, thus they are nonstationary. In this paper we concentrate on the nonstationary case, and propose a novel fast iterative technique for real-time tracking of the time-varying frequencies. Simulation examples are given which compare the iterative method to other data-adaptive techniques. These examples show that the proposed method converges to good estimates of the frequencies much more quickly than the adaptive techniques as well as being able to perform better in tracking nonstationary sinusoids.

In multipulse LPC speech coding, there are a number of important issues which need to be addressed. The purpose of this paper is to report some of our work related to two of these issues. The first pertains to how one measures the quality of the synthetic speech for a given set of multipulse parameters. In this paper, we present a general set of frequency domain weighting functions which may incorporate various spectral characterizations of the speech. The second issue relates to how one determines the optimal set of pulse positions for a given set of of LPC filter coefficients. Current methods involve a sequential procedure, which is suboptimal compared with simultaneous positioning of pulses throughout an entire analysis frame. We therefore discuss an algorithm for simultaneous pulse positioning based on Pisarenko's harmonic decomposition.

Many techniques are available for the estimation of the power spectrum of a stationary random process. While power spectrum estimation is a problem which falls within the domain of signal processing, the problem of inferring information falls within the domain of artificial intelligence (AI). With a wide variety of different types of power spectrum estimation techniques to choose from, an equally wide range of differing spectral estimates may be produced. Each estimate, however, may be used to infer information about the time series. By defining an appropriate knowledge base, a system is being developed to infer information from power spectrum estimates. This system combines the estimates produced by a variety of current spectrum estimation techniques in order to formulate a composite spectral estimate.

In this paper a class of iterative procedures is presented for reconstructing a finite duration sequence from noisy samples of its Fourier transform phase. These measurements are combined with a knowledge of the true transform magnitude and/or hard constraints on the phase noise variations to define sets whose intersection must contain the true sequence. The algorithms iterate between the sequence domain and the transform domain applying the known constraints (i.e. finite duration and known limits on phase variation) in each domain. Results of an experimental investigation are presented. A method is described for the case where limits on both the magnitude and phase variation of a finite length sequence are known.

In this paper, the importance of spectral phase and magnitude is examined from a different point of view. In particular, an amplitude and angle based representation of spectral information is developed. With this formulation, a causal finite length signal is uniquely defined by its spectral amplitude or, to within a scale factor, by its spectral angle.

The extrapolation of a bandlimited signal from observations taken over a finite interval is an important problem in applications such as spectral estimation and image processing. The extrapolation of bandlimited analog signals is fundamentally different, however, from the band-limited extrapolation of discrete time sequences. Specifically, as has been previously noted, a unique solution to the discrete problem does not exist. In this paper, this fact is demonstrated in a new and convincing way. In particular, two methods are presented for constructing a sequence which, in the frequency domain, is band-limited to an arbitrary cutoff frequency and which, in the time domain, is equal to zero over an interval of arbitrary length. The importance of the existence of these multiple solutions to the general extrapolation problem is emphasized and questions are raised regarding the need for additional constraints in the discrete band-limited extrapolation problem.

In this paper, the importance of the boundary values of a two dimensional sequence in the phase retrieval problem is investigated. Specifically, it is shown that, given the boundary values, phase retrieval becomes a linear problem. Therefore, the interior points of the two-dimensional sequence may easily be deduced from the boundary values of the sequence and the magnitude of its Fourier transform or, equivalently, the autocorrelation function of the sequence. Furthermore, although the determination of the boundary values from only Fourier transform magnitude information is, in general, a non-trivial problem, it is shown that the shape of the region of support of the two-dimensional sequence determines the ease with which the boundary values may be determined.

This paper reviews some recent results concerning the uniqueness of a multi-dimensional (m-D) sequence in terms of the phase or magnitude of its Fourier transform. Specifically, some conditions are presented under which an m-D sequence is uniquely specified by its phase or magnitude. In addition, an iterative algorithm is described which has been proposed for reconstructing a sequence from its phase or magnitude. Several examples of phase-only and magnitude-only reconstruction are also presented.

This paper is concerned with the development of a general convergence proof which is applicable to a particular class of iterative signal reconstruction algorithms. The proof relies on the concept of a nonexpansive mapping and the uniqueness of the desired signal. Two specific iterative reconstruction algorithms which are considered to illustrate this method of proof are bandlimited extrapolation and phase-only signal reconstruction, the convergence proof of the later being a new result. The generality of the approach allows for the incorporation of nonlinear constraints such as positivity or minimum and maximum value constraints.

In this paper, we develop a set of conditions under which a sequence is uniquely specified by the phase or samples of the phase of its Fourier transform. These conditions are applicable to mixed-phase one-dimensional and multi-dimensional sequences. Under the specified conditions, we also present several algorithms which may be used to reconstruct a sequence from its phase.