Image segmentation is a key task in computer vision and image processing with important applications such as scene understanding, medical image analysis, robotic perception, video surveillance, augmented reality, and image compression, among others, and numerous segmentation algorithms are found in the literature. Against this backdrop, the broad success of deep learning (DL) has prompted the development of new image segmentation approaches leveraging DL models. We provide a comprehensive review of this recent literature, covering the spectrum of pioneering efforts in semantic and instance segmentation, including convolutional pixel-labeling networks, encoder-decoder architectures, multiscale and pyramid-based approaches, recurrent networks, visual attention models, and generative models in adversarial settings. We investigate the relationships, strengths, and challenges of these DL-based segmentation models, examine the widely used datasets, compare performances, and discuss promising research directions.

This paper presents combining two data augmentation methods involving speed perturbation and room impulse response reverberation for the purpose of improving the generalization capability of convolutional neural networks when used for voice command recognition. Speed perturbation generates voice command variations caused by shorter or longer time durations of commands spoken by different speakers. Room impulse response reverberation generates voice command variations caused by reflected sound paths. The combination of these two augmentation methods is presented in this paper by examining a public domain dataset of voice commands. The experimental results based on the performance metric of word error rate indicate the improvement in voice command recognition rates when combining these data augmentation methods relative to using each augmentation method individually.

This paper presents an educational tool to learn about how hearing aid compression fitting is prescribed from a signal processing perspective. An interactive web-based program has been developed based on the widely used DSL-v5 fitting rationale. This program can be accessed and used from any internet browser to generate the parameters of compression curves that correspond to the nine frequency bands used in DSL-v5. These parameters are then transferred to a smartphone in the form of a datafile to be used by a compression app, which operates as a virtual hearing aid, running in real-time on both iOS and Android smartphones. This educational tool was found to be very easy-to-use by signal processing engineers as no programming knowledge is needed for adjusting gain across frequency bands and subsequently running the corresponding compression curves in the smartphone app to appropriately compress input sound signals.

This paper presents a multi-label convolutional neural network approach to determine the number of speakers when using a single microphone which is more challenging than when using multiple microphones. Spectrograms of windowed noisy speech signals for 1talker, 2talkers and 3+talkers are used as inputs to a multi-label convolutional neural network. The architecture of the developed multi-label convolutional neural network is discussed and it is shown that this network with median filtering can achieve an overall accuracy of about 81% for the noisy speech dataset examined.

This paper presents a newly developed laboratory paradigm for the laboratory component of signals and systems courses. It involves the implementation of signals and systems algorithms that are written in MATLAB on smartphones using their ARM processors. This smartphone-based approach enables an anywhere-anytime environment for students to conduct signals and systems experiments. The steps involved for running MATLAB codes on smartphones are discussed together with the laboratory experiments that are normally covered in signals and systems courses. It is shown that this paradigm still keeps the programming language for conducting signals and systems experiments to MATLAB while enabling a truly mobile laboratory environment for students to learn the implementation aspects of signals and systems concepts.

This paper presents a human action recognition approach using a depth camera for situations when actions of interest are performed in a continuous and random manner among actions of non-interest. The developed approach first performs detection of actions of interest by separating actions of interest from actions of non-interest in an on-the-fly manner and then classifies the detected actions of interest. Skeleton joint positions from depth images are used to achieve the detection of actions of interest. Recognition of detected actions of interest is then achieved by fusing the outcome of two classifiers, one classifier using skeleton joint positions and the other classifier using depth images. A continuous dataset consisting of actions of interest associated with the smart TV application is collected and made publicly available. The results obtained by applying the developed approach to this dataset indicate its effectiveness in detecting and recognizing actions of interest from continuous data streams.

This paper presents a real-time unsupervised classification of environmental noise signals without knowing the number of noise classes or clusters. A previously developed online frame-based clustering algorithm is modified by adding feature extraction, a smoothing step and a fading step. The developed unsupervised classification or clustering is examined in terms of purity of clusters and normalized mutual information. The results obtained for actual noise signals exhibit the effectiveness of the introduced unsupervised classification in terms of both classification outcome and computational efficiency.

This paper presents a human action recognition approach by the simultaneous deployment of a second generation Kinect depth sensor and a wearable inertial sensor. Three data modalities consisting of depth images, skeleton joint positions, and inertial signals are fused by utilizing three collaborative representation classifiers. A database consisting of 10 actions performed by 6 subjects is put together to carry out two types of testing of the developed fusion approach: subject-generic and subject-specific. The overall recognition rates obtained from both types of testing indicate recognition improvements when fusing all the data modalities compared to the situations when data modalities are used individually.

This paper presents the real-time implementation and field testing of an app running on smartphones for classifying noise signals involving subband features and a random forest classifier. This app is compared to a previously developed app utilizing mel-frequency cepstral coefficients features and a Gaussian mixture model classifier. The real-time implementation has been carried out on both the Android and iOS smartphones. The field testing results indicate the superiority of this newly developed app over the previously developed app in terms of classification rates.

This paper presents the steps one needs to take in order to run a signal processing algorithm designed in Simulink on the ARM processor of smartphones. The steps are conveyed by transitioning two signal processing application examples from Simulink to smartphone. The application examples involve background noise classification and lane departure detection. Considering that Simulink programming is widely used in signal processing, the approach presented in this paper is of benefit to practicing engineers and signal processing researchers/educators in terms of the process of making Simulink codes or models to run on smartphones.

Human action recognition has a wide range of applications including biometrics, surveillance, and human computer interaction. The use of multimodal sensors for human action recognition is steadily increasing. However, there are limited publicly available datasets where depth camera and inertial sensor data are captured at the same time. This paper describes a freely available dataset, named UTD-MHAD, which consists of four temporally synchronized data modalities. These modalities include RGB videos, depth videos, skeleton positions, and inertial signals from a Kinect camera and a wearable inertial sensor for a comprehensive set of 27 human actions. Experimental results are provided to show how this database can be used to study fusion approaches that involve using both depth camera data and inertial sensor data. This public domain dataset is of benefit to multimodality research activities being conducted for human action recognition by various research groups.

Exposure bracketing is widely used to generate high dynamic range (HDR) images. Exposure bracketing often consists of an image taken by the auto-exposure setting of a camera as well as two other images taken at ±n exposure levels away from the auto-exposure setting where n is manually selected or specified. The images in the bracket are then blended to generate an HDR image. This paper presents an automatic exposure selection method that relieves the user from setting the exposure level, thus allowing HDR images to be generated with no user intervention. This is achieved by using the camera characteristics function and the scene information. The results obtained indicate the HDR images generated by using this method outperform the manual selection of exposure level in terms of two widely used image quality measures.

This paper presents a computationally efficient method for action recognition from depth video sequences. It employs the so called depth motion maps (DMMs) from three projection views (front, side and top) to capture motion cues and uses local binary patterns (LBPs) to gain a compact feature representation. Two types of fusion consisting of feature-level fusion and decision-level fusion are considered. In the feature-level fusion, LBP features from three DMMs are merged before classification while in the decision-level fusion, a soft decision-fusion rule is used to combine the classification outcomes. The introduced method is evaluated on two standard datasets and is also compared with the existing methods. The results indicate that it outperforms the existing methods and is able to process depth video sequences in real-time.

This paper presents a fusion approach for improving human action recognition based on two differing modality sensors consisting of a depth camera and an inertial body sensor. Computationally efficient action features are extracted from depth images provided by the depth camera and from accelerometer signals provided by the inertial body sensor. These features consist of depth motion maps and statistical signal attributes. For action recognition, both feature-level fusion and decision-level fusion are examined by using a collaborative representation classifier. In the feature-level fusion, features generated from the two differing modality sensors are merged before classification, while in the decision-level fusion, the Dempster-Shafer theory is used to combine the classification outcomes from two classifiers, each corresponding to one sensor. The introduced fusion framework is evaluated using the Berkeley multimodal human action database. The results indicate that because of the complementary aspect of the data from these sensors, the introduced fusion approaches lead to 2% to 23% recognition rate improvements depending on the action over the situations when each sensor is used individually.

Howling is one of the most annoying consequences of the acoustical coupling in hearing aids. In this paper, a method of howling detection is proposed applying Discrete Energy Separation Algorithm-2 (DESA-2) and Generalized Teager-Kaiser Operator (GTKO). Several GTKOs with various resolution parameters monitor the input signal to recognize the howling occurrence while DESA-2 is used to estimate the howling frequency. Performance of the proposed method is compared with two known howling detection approaches, i.e. Peak-to-Harmonic Power Ratio (PHPR) and Adaptive Notch Filter (ANF). Simulation results show the advantage of the proposed method in terms of having lower false detection and shorter detection time.

This paper presents improvements made to the previously developed noise classification path of the environment-adaptive cochlear implant speech processing pipeline. These improvements consist of the utilization of subband noise features together with a random forest tree classifier. Three commonly encountered noise environments of babble, street, and machinery are considered. The results using actual noise signals indicate that this classification method provides 10% improvement in the overall classification rate compared to the previously developed classification while maintaining the real-time implementation aspect of the entire speech processing pipeline.

This paper examines environment-adaptive noise suppression algorithms for computationally efficient or real-time implementation in bilateral cochlear implants using a single processor. A generalized framework is introduced that allows one to train suppression and head-related transfer function gain tables not only for different noise environments but also for different distortion measures. This generalization incorporates any differentiable measure with unilateral data-driven enhancement methods becoming its special cases. Specifically, the solutions for three distortion measures of Weighted-Euclidean, Log-Euclidean and Weighted-Cosh are provided. These solutions are evaluated in six commonly encountered noise environments for a wide range of directionalities.

This paper presents our recent development of a classification algorithm for identification of breast cancer margins measured by hyperspectral imaging for the purpose of lowering the number of missed positive margins in breast cancer lumpectomy. After extracting Fourier coefficient selection features and reducing the dimensionality of hyperspectral image data via the Minimum Redundancy Maximum Relevance method, an SVM classifier involving a radial basis kernel function is deployed to separate cancerous tissues from normal tissues. By examining exvivo breast cancer hyperspectral images tagged by a pathologist, the developed classification approach is shown to achieve a sensitivity of about 98% and a specificity of about 99%.

The Levenberg-Marquardt optimization is normally used in 3D point cloud reconstruction from image sequences which is computationally expensive. This paper presents a two-stage camera pose estimation approach where an initial camera pose is obtained during the first stage and a refinement is performed during the second stage. This approach does not require the use of the Levenberg-Marquardt optimization and LU matrix decomposition for computing the projection matrix, thus providing a more computationally efficient 3D point cloud reconstruction as compared to the existing approaches. The results obtained using real video sequences indicate that the introduced approach generates lower re-projection errors as well as faster 3D point cloud reconstruction.

In this paper, the properties of a new class of causal Markov random fields, named symmetric Markov mesh random field, are initially discussed. It is shown that the symmetric Markov mesh random fields from the upper corners are equivalent to the symmetric Markov mesh random fields from the lower corners. Based on this new random field, a symmetric, corner-independent, and isotropic image model is then derived which incorporates the dependency of a pixel on all its neighbors. The introduced image model comprises the product of several local 1D density and 2D joint density functions of pixels in an image thus making it computationally tractable and practically feasible by allowing the use of histogram and joint histogram approximations to estimate the model parameters. An image restoration application is also presented to confirm the effectiveness of the model developed. The experimental results demonstrate that this new model provides an improved tool for image modeling purposes compared to the conventional Markov random field models.

Applied digital signal processing courses offered at many universities do not normally include FPGA implementation of signal processing algorithms. This is due to the fact that students taking applied signal processing courses often are not familiar with FPGA programming languages and as a result FPGA implementation cannot be easily achieved in one-semester long applied digital signal processing courses. In this paper, we present an approach for making FPGA implementation easily achievable within one-semester long applied digital signal processing courses or within a relatively short design cycle time. Our approach is based on the LabVIEW FPGA Module which allows FPGA implementation to be achieved in a time-efficient manner via more intuitive graphical coding. An overview of our designed FPGA implementation approach covering typical laboratories that students encounter in applied digital signal processing courses is provided.

Stochastic image modeling based on conventional Markov random fields is extensively discussed in the literature. A new stochastic image model based on Markov random fields is introduced in this paper which overcomes the shortcomings of the conventional models easing the computation of the joint density function of images. As an application, this model is used to generate texture patterns. The lower computational complexity and easily controllable parameters of the model makes it a more useful model as compared to the conventional Markov random field-based models.

Auto-focusing is a key feature of modern camera systems that ensures a focused image is captured without the need for any user intervention. Passive auto-focusing is achieved based on a measure of sharpness or a sharpness function that is extracted from a series of images at different focal positions. With the rapid growth of stereo camera systems for 3D applications, this paper introduces a sharpness function when a series of stereo image pairs are captured at different focal positions. This function is defined based on the joint information between the left and right images. A comparison is carried out between the introduced sharpness function and the conventional sharpness functions when considering the left and right images separately in terms of sharpness accuracy and computation time.

This paper presents the real-time implementation of wavelet-based Advanced Combination Encoder on PDA platforms for cochlear implant studies. Three real-time implementations using the conventional FFT, our previous recursive DFT, and the wavelet transform are compared in terms of computational complexity, processing time and fixed-point accuracy. The results obtained show that this new real-time implementation on PDA platforms is computationally comparable to our previous real-time recursive DFT implementation while achieving a higher accuracy or a lower quantization error.

This paper presents an image processing system for detecting mouse preterm labor using Second Harmonic Generation (SHG) microscopy images. Two classes of SHG images are considered: normal pregnant cervix and premature cervical remodeling induced by Mifepristone. Among the commonly used texture features in image processing, wavelet-based texture features together with previously utilized image features for SHG microscopy of artificial collagen gels are identified to form an effective set of features for distinguishing the two classes of images. The results obtained indicate that correct detection rates above 98% are achievable.

This paper presents a new framework for facial expression recognition based on diffeomorphic matching. First landmarks are selected based on a manual or automatic method. All of the landmarks from different images are registered to a reference landmark set using a rigid registration algorithm. The pair-wise geodesic distance between all sets of landmarks are then computed using diffeomorphic matching. Finally, a K-Nearest Neighbor classifier (KNN) is used to classify a query image using the geodesic distances. Both the classification and classical MultiDimensional Scaling results show that geodesic distance is more effective than Euclidean distance on capturing the face shape variation.

This paper presents a real-time and interactive implementation of the recursive Fourier transform approach on PDA platforms for cochlear implant signal processing applications. PDA platforms provide a cost-effective and portable platform for cochlear implant studies. The computational complexities of the two commonly used signal processing strategies that are being used in commercial cochlear implants are compared. A more computationally efficient approach using recursive Fourier transform is discussed and a real-time implementation of this approach is then accomplished on a PDA platform. Different versions of the implementation are examined and compared in terms of speed and accuracy.

Although many face detection algorithms have been introduced in the literature, only a handful of them can meet the real-time constraints of mobile devices. This paper presents the real-time implementation of our previously introduced face detection algorithm on a mobile device. The steps taken to achieve such a real-time implementation are discussed. Real-time comparison results with the widely used Viola-Jones face detection algorithm in terms of detection rate and processing speed are presented to demonstrate the robustness of our real-time solution.

In low-light conditions, the focus value curve or sharpness function that is used in passive auto-focusing becomes flat due to the increase in the noise level. This paper presents an enhanced low-light auto-focusing system model which provides performance improvements over our previously introduced low-light auto-focusing model. This is achieved by utilizing an indexed family of focus value curves via a digital band-pass filterbank. A new sharpness ratio metric is presented for evaluation of auto-focusing in low-light conditions. Experimental results obtained from three different prototype camera platforms demonstrate the improved low-light performance over our previous model.

Although there are many face detection algorithms in the literature, only a handful of them meet the real-time constraints of a software-based solution without using any dedicated hardware engine. This paper presents a real-time and robust solution for mobile platforms which in general have limited computation and memory resources as compared to PC platforms. This solution involves combining our two previous real-time implementations for mobile platforms to address the shortcoming of each implementation. The first implementation provides an online or on-the-fly light source calibration for the second implementation which is found to be robust to various face poses or orientations. The real-time results obtained on an actual mobile platform indicate both the real-time and robustness capabilities of this hybrid face detection solution.

Image registration is a key component in existing magnetic resonance image processing software packages. Because of nonlinear variability in brain substructures, the commonly used registration pipelines applied to the entire brain area may not lead to accurate registration for substructures. This paper presents a two-stage registration approach for a better alignment of brain substructures or regions-of-interest. In the first stage, an affine transformation function is applied to the entire brain area and in the second stage, a nonrigid or deformable transformation function is applied to the substructure of interest. To show the usefulness of this two-stage registration, images from the Brainweb database are examined using normalized mutual information is used to assess the degree of alignment. The comparison results indicate improvements over the commonly used registration approaches.

Shannon entropy is a powerful tool in image analysis, but its reliable computation from image data faces an inherent dimensionality problem that calls for a low-dimensional and closed form model for the pixel value distributions. The most promising such models are Markovian, however, the conventional Markov random field is hampered by noncausality and its causal versions are also not free of difficulties. For example, the Markov mesh random field has its own limitations due to the strong diagonal dependency in its local neighboring system. A new model, named quadrilateral Markov random field (QMRF) is introduced in this paper in order to overcome these limitations. A property of QMRF with neighboring size of 2 is then used to decompose an image prior into a product of 2-D joint pdfs in which they are estimated using a joint histogram under the homogeneity assumption. In addition, the paper includes an extension of the introduced method to the computation of image spatial mutual information. Comparisons on synthesized images as well as two applications with real images are presented to motivate the developments in this paper and demonstrate the advantages in the performance of the introduced method over the existing ones.

Signals and systems lab courses that are currently offered at many universities are mostly based on text-based programming languages and environments, in particular MATLAB. This paper presents an alternative programming approach in these courses by combining textual with graphical programming. This hybrid programming approach offers code interactivity in a time-efficient manner. Example labs are presented to demonstrate how this approach allows students to interact and experiment with their codes to gain a better understanding of the signal processing concepts.

The accuracy and computational complexity of mutual information (MI) estimation are critical factors in multi-modality non-rigid image registration. This paper discusses the accuracy and complexity of MI estimation approaches based on non-rigid registration functions. General formulations have been derived for Shannon's and Renyi's definitions of MI, as well as Cauchy- Schwartz quadratic MI. The results obtained indicate that a fuzzy histogram binning estimation approach is significantly faster and more accurate than the conventional non-parametric Parzen window estimation approach. The analytical formulations obtained for various MI definitions are continuously differentiable and are shown to be computationally efficient for high-dimensional optimization problems particularly for non-rigid image registration.

In the scalable video coding extension of the H.264/AVC standard, an additional inter-layer prediction is combined with the traditional inter-frame prediction and intra prediction to obtain coding efficiency. This additional prediction demands a higher video encoding computational complexity. In this paper, our previously developed fast adaptive early termination algorithm is utilized to speed up the mode selection component of the H.264 scalable video coding at the base and enhancement layers. The experimental results over different content video sequences show a time reduction by 20% to 50% for spatial and quality scalable situations as compared with JSVM9.8 while achieving more or less the same video quality.

This paper presents the real-time implementation of a cochlear implant signal processing system on PDA platforms. PDAs are chosen as they provide portable and cost-effective computation platforms. To gain software flexibility and interactivity, the LabVIEW graphical programming environment is used. The paper discusses the optimization steps which are taken to achieve a real-time throughput. These steps consist of using dynamic link libraries, utilizing efficient memory allocation, and performing fixed-point arithmetic. These steps are general purpose in the sense that the same steps can be deployed for real-time implementation of other clinical or industrial signal processing applications on PDAs.

This paper presents the use of Kullback-Leibler Distance (KLD) as part of an optimization framework to incorporate prior knowledge from field maps into non-rigid registration of echo-planar (EPI) to structural magnetic resonance brain images. An analytical expression is derived for the derivatives of KLD with respect to registration transformation parameters, which is shown to be computationally more efficient as compared to the derivatives of mutual information. Quantitative gold standard validation is carried out on simulated digital brain phantom images with synthesized deformations. In addition, in-vivo validation is performed via a cross-comparison of the similarity of high-resolution and low-resolution EPI to T1-and T2-weighted structural images. The results obtained indicate that the developed KLD-based non-rigid registration technique provides an effective way of correcting local distortions in echo-planar imaging.

Image blur occurs often when using a cell-phone camera due to handshakes or jitter. Although there exist many motion deblurring algorithms in the literature, the computational complexities of these algorithms and the assumptions considered make them unsuitable for deployment on a cell-phone camera processor. This paper presents an image blur reduction algorithm for cell-phone cameras having a low computational complexity and without making any assumption about the handshake motion. This algorithm utilizes one low-exposure image in addition to a blurred image to perform a blur reduction operation via tonal correction. The developed tonal correction approach is adaptive to the scene by taking into consideration the brightness and contrast of the blurred image. The results obtained indicate the effectiveness of this blur reduction algorithm for handshake blurred images captured by a cell-phone camera.

Functional localization is a concept which involves the application of a sequence of geometrical and statistical image processing operations in order to define the location of brain activity or to produce functional/parametric maps with respect to the brain structure or anatomy. Considering that functional brain images do not normally convey detailed structural information and, thus, do not present an anatomically specific localization of functional activity, various image registration techniques are introduced in the literature for the purpose of mapping functional activity into an anatomical image or a brain atlas. The problems addressed by these techniques differ depending on the application and the type of analysis, i.e., single-subject versus group analysis. Functional to anatomical brain image registration is the core part of functional localization in most applications and is accompanied by intersubject and subject-to-atlas registration for group analysis studies. Cortical surface registration and automatic brain labeling are some of the other tools towards establishing a fully automatic functional localization procedure. While several previous survey papers have reviewed and classified general-purpose medical image registration techniques, this paper provides an overview of brain functional localization along with a survey and classification of the image registration techniques related to this problem.

LabVIEW and Simulink are two most widely used graphical programming environments for designing digital signal processing (DSP) systems. Unlike conventional text-based programming languages, such as C and MATLAB, graphical programming involves block-based code development, allowing a more efficient mechanism to build and analyze DSP systems. This paper presents a comparative evaluation between LabVIEW and Simulink in terms of a number of ease-of-use and functionality criteria. Twenty students taking a senior undergraduate DSP lab course were asked to perform the evaluation. The students' responses indicate that these two graphical environments provide more or less the same design features with LabVIEW having an edge over Simulink as far as graphical display/visualization and DSP hardware integration tools are concerned

Functional to anatomical brain image registration is needed for accurate localization of brain activation maps. Due to the presence of nonlinear distortions, it is more effective to consider non-rigid transformations to achieve such a registration. In this paper, a non-rigid registration technique based on the B-spline free-form deformation model and mutual information similarity measure is introduced. An optimization formulation is devised to achieve a fast and robust registration. This formulation differs from the previous formulations by utilizing a limited-memory, second-order optimization algorithm instead of the usual first-order gradient-based algorithms. It also enforces hard parameter constraints instead of constraints based upon physics or Jacobian smoothness. The results obtained indicate that this registration technique provides improvements over rigid and affine techniques when registering functional to anatomical magnetic resonance brain images.

This paper addresses the problem of classification of electroencephalogram (EEG) signals obtained from human subjects performing two mental tasks. One task named baseline involves relaxing and thinking of nothing in particular and the other task named multiplication involves mentally multiplying two 2-digit integers. First, the EEG signals are pre-processed using independent component analysis for removal of artifacts. Then, a time-frequency representation of the signals is generated, from which wavelet-based texture features are extracted for classification. The texture features are fed into a three-layer neural network classifier trained by the backpropagation algorithm. A classification rate of 96% is obtained for the dataset examined. The entire classification system has been implemented in the LabVIEW graphical programming environment providing a user-friendly interface to alter and monitor various parameters of the neural network classifier.

This paper discusses a real-time digital signal processor (DSP)-based hierarchical neural network classifier capable of classifying both analog and digital modulation signals. A high-performance DSP processor, namely the TMS320C6701, is utilized to implement different kinds of classifiers including a hierarchical neural network classifier. A total of 31 statistical signal features are extracted and used to classify 11 modulation signals plus white noise. The modulation signals include CW, AM, FM, SSB, FSK2, FSK4, PSK2, PSK4, OOK, QAM16, and QAM32. A classification hierarchy is introduced and the genetic algorithm is employed to obtain the most effective set of features at each level of the hierarchy. The classification results and the number of operations on the DSP processor indicate the effectiveness of the introduced hierarchical neural network classifier in terms of both classification rate and processing time.

Individual identification of dolphins is of much interest to marine mammologists. Natural markings and notches on the trailing edge of the dorsal fin provide an almost unique signature within a study population for individual recognition. The dorsal fin is an almost rigid, planar object, and, as a result, the problem boils down to planar curve matching. Invariants under three groups of similarity, affine and projective transformations are examined in this paper. Three families of differential, algebraic and semi-differential invariants are studied and evaluated for their applicability and performance in dolphin identification from photographs. It is observed that, under practical conditions, for smooth quasi-rigid curves like the trailing edge of the dorsal fin, curvature-based matching methods outperform almost all of the other techniques which offer a higher level of invariance.

Elastography is a new imaging modality which provides an image of tissue elastic properties. Simulated elastograms are used to study the ability of textural features to discriminate between abnormalities with different levels of hardness and density. Elastograms are simulated for different signal processing parameters as well as for a non-uniform stress field. 15 selected texture features are extracted. The Gaussian classifier does not perform well due to the complexity of clusters. Hence, a neural network classifier is trained and tested to discriminate between the classes. The results indicate that the texture features perform well under variations of the signal processing parameters and non-uniformity of the applied stress field. This simulation study provides a guiding mechanism to discriminate between malignant and benign tissue abnormalities once clinical data becomes available.

This paper presents a scheme for the classification of mass abnormalities in digitized or digital mammograms based on two novel image texture features. The first texture feature provides a measure of smoothness/denseness and is obtained by applying a morphological operator to maxima/minima image points. The second texture feature reflects a measure of architectural distortion and is derived from image gradients. A three-layer backpropagation neural network is used as the classifier. The performance of the classification scheme is evaluated by carrying out a receiver operating characteristic (ROC) analysis. Classification of 150 biopsy proven masses into benign and malignant classes resulted in a ROC area of 0.91. The results obtained demonstrate the potential of using this scheme as an electronic second opinion to lower the number of unnecessary biopsies.

A novel approach is presented for characterizing transient wandering tones. These signals are segmented and approximated as time series with piecewise linear instantaneous frequency and piecewise constant amplitude. Frequency rate, center frequency, and energy features are estimated in each segment of data using chirped autocorrelations and the fractional Fourier transform. These features are tracked across segments using linear dynamical models whose parameters are estimated using an expectation-maximization algorithm. A new cross-covariance estimator for adjacent states of the dynamical model is given. The feature extraction/tracking algorithm is used to characterize a measured marine-mammal vocalization. Application of the representation algorithm to signal classification is discussed.

Elastography is an upcoming and promising imaging modality which is used to map as an image, termed elastogram, information related to tissue elastic properties. An elastogram reflects strain information resulting from an applied stress. This paper presents a study of various image texture features as applied to simulated elastograms for the purpose of obtaining effective texture features. These features are those that exhibit non-overlapping distributions for different modulus distributions. Among various texture features studied, five are found to be effective for simulated elastograms. These include fractal signature, wavelet energy, and entropy, contrast and energy extracted from the sum and difference histograms. Upon the availability of clinical data, the results of this study are used to classify abnormalities into malignant or benign tissue classes.

This paper describes the application of signal processing methods to extract key features for detection and prediction of faults in rotating mechanical components of a whirl tower test facility. Procedures are described for processing the vibration signals from critical components of the whirl tower using the Fourier transform and the wavelet transform. The wavelet transform provides localization of signals in both time and frequency, revealing data that is averaged out in the Fourier analysis. The features extraction processes described are based on specific knowledge of the whirl tower equipment. The real-time analysis discussed allow for scheduling of inspection, repair, or replacement of failed and degraded components with minimal impact on production.

Presents a comprehensive discussion on the segmentation of mammograms using morphological texture features. These features are derived from morphological granulometries with various structuring elements. Each structuring element captures a specific texture content. The segmentation is carried out in an unsupervised manner by applying the KL (Karhunen-Loeve) transform feature reduction and Voronoi clustering on the extracted morphological texture features. The evaluation of the segmentation outcome by a trained radiologist is provided.

The presence of microcalcifications in mammograms provides an early indication of possible breast cancer. Because of the difficulty associated with visual identification of microcalcifications and the large volume of mammograms read per day, the radiologist stands a good chance of missing some small microcalcification clusters. Although several computer-assisted programs have been developed for the automatic detection of microcalcifications in mammograms, they often generate too many false positives. This paper presents a computer-assisted enhancement technique which is capable of coping with false positive samples. More specifically, a general-purpose clustering algorithm, called Issac (Interactive Selective and Adaptive Clustering), has been developed which achieves a compromise between sensitivity and generalization attributes of existing clustering algorithms. Issac comprises two parts: (i) selective clustering and (ii) interactive adaptation. The first part reduces the number of false positives by identifying sensitive sample domains in the feature space. The second part allows the radiologist to improve results by interactively identifying additional false positive or true negative samples. The clinical evaluation of the results has indicated that the developed enhancement technique has the potential of being an effective mechanism to bring microcalcification areas to the attention of the radiologist during a routine reading session of mammograms. Further clinical evaluation is being carried out for the purpose of full-scale clinical deployment.>

When a camera is used to provide the navigational parameters in autonomous vehicle operations, it is subjected to unexpected movements or vibrations of the mounting platform. This paper presents a framework for analyzing the effect of uncontrollable camera movements on the navigational parameters, in particular on the range and-heading angle in vision-based vehicle tracking. The noise introduced by the platform movements is modeled in two ways: camera noise approach and image noise approach. The parameter space of the camera is divided into a controllable subspace consisting of its height and depression angle, and an uncontrollable subspace consisting of the tracked object coordinates and rotation angle errors. A consistent detectable region is then obtained such that the tracked object is always seen by the camera. Based on this region, a reliable region consisting of no singularity points is established so that the range error does not become infinity. The optimum parameters of the controllable subspace with respect to the uncontrollable subspace are found by employing two estimation schemes: (a) the mini-max estimator to provide the worst case effect, and (b) the minimum-mean-square estimator to provide the average or overall effect. From the results obtained, it is shown how an optimum imaging geometry of a monocular vision-based tracking system can be designed in order to satisfy prescribed levels of range and heading angle errors.>

In an outdoor navigational environment, cameras of an autonomous vehicle or robot are subjected to uncontrollable movements relative to the vehicle or robot body. In this paper, the effect of such camera movements in stereo vision-based vehicle tracking systems is analyzed. It is shown how the computation of the range and heading angle of a lead vehicle is affected as a result of camera movements. The parameter space of the stereo system is divided into a controllable subspace consisting of the baseline and an uncontrollable subspace consisting of relative camera movements and the coordinates of a tracked point on the lead vehicle. The optimal baseline value is found by employing the mini-max and minimum mean-squared error estimators. The mini-max estimator is used to obtain the worst case performance while the minimum mean-squared estimator to obtain the average performance. Based on this analysis, it is shown how an optimal imaging geometry for a stereo vision-based tracking system can be designed in order to have the least amount of range and heading angle errors.>

In the technical discipline of computer vision, the concept of using stereo cameras for depth perception has been motivated by the fact that, in human vision, one percept can arise from two retinal images as a result of the process of “fusion” in the visual cortex. If mankind wants to apply this concept to computer controlled machines it must be inferred that knowledge of the psychological process is sufficient to emulate, at least in a weak sense, the human sensory phenomena called vision. Typically, models for this emulation process have been computationally intensive. They contained or required massive amounts of data and thus resulted in very time consuming analytic solutions for depth or range information. The visual process, however complex, does not operate independently of other brain functions. For instance, the superior colliculus of the brain is used to generate information about visual field motion and many visual reflexes. The superior colliculus is also a multi-modal sensory area which includes auditory and somaethetic information [1]. This visual system-brain function connection must be kept in mind when one is attempting to utilize general visual system models to control manipulative processes in order to accomplish a task. In this paper we investigate the utilization of binocular fusion to determine range and heading angle for the specialized control task of guidance for an autonomous vehicle in a convoy following application. It is therefore necessary to develop the vision model, to define the decision space and to use multi-modal information when appropriate. Both parallel and convergent vision models have been utilized for this task. A comparison of calibration results, the mathematical advantages of a log-polar decision space and the resulting real-time structure for controlling a vehicle in a convoy mode are given from actual experimental evaluation.

A collision-free navigation scheme for an autonomous vehicle is presented. An attempt is made toward establishing a more realistic navigational environment by considering moving obstacles with unknown trajectories. A collision-free path between the current location and a desired location of the vehicle is obtained, based on future positions of the obstacles. The future positions are predicted using an autoregressive (AR) model. A simulation example is implemented where this scheme is applied to a mobile robot in the presence of two moving objects.>

The Darboux vector contains both the curvature and torsion of a three-dimensional (3-D) image as its components. Curvature measures how sharply a curve is turning while torsion measures the extent of its twist in 3-D space. Curvature and torsion completely define the shape of a 3-D curve. A scheme is presented that uses the length of this vector, also called the total curvature, for the segmentation of 3-D contours. A quintic B-spline is used in this formulation to obtain the total curvature for noisy data. Examples of nonnoisy and noisy data illustrate the merit of the scheme.>

This paper presents a new scheme for segmentation of three-dimensional (3D) contours based on a criterion involving their curvature and torsion properties. Curvature indicates how sharply a curve is turning while torsion measures the extent of its twist in 3D space. Also, curvature and torsion completely define a curve in 3D space. For these reasons, the two together have been chosen here as a set of intrinsic and complete structural features of a contour. The Darboux vector Ω contains both these features as its components. We use the length ∥Ω∥ of the Darboux vector, also known as total curvature, as the criterion for segmentation of the contour. Specifically we choose the zero-crossings of d2∥Ω∥/ds2, where s = arc length, as the breakpoints in the segmentation of the contour. The segmentation scheme is described by means of a block diagram which includes filters for calculation of d2∥Ω∥/ds2in terms of appropriate distributional derivatives resulting from the contour parametric equations.