Computer-aided Gleason grading of prostate cancer tissue images has been in rapid development during the past decade. Automated classifiers using features derived from multiwavelet transform, fractal dimension and other measurements, and using texton forests have shown considerable successes. This paper presents our study on application of cardinal multiridgelet transform (CMRT) to prostate cancer images to extract texture features in the transform domain. CMRT can provide cardinality, approximate translation invariance and rotation invariance simultaneously. With 32 images of Gleason grade 3 and grade 4 as a training set and using texture features extracted therefrom, a support vector machine with Gaussian kernel has been trained to classify grade 3 and grade 4. The leave-one-out cross-validation showed its accuracy of 93.75% and AUC of 0.9651. 10 test images of grade 4 showed 100% accuracy.

This paper describes an algorithm to decompose speech into tonal, transient, and residual components. The algorithm uses an MDCT-based hidden Markov chain model to isolate the tonal component and a wavelet-based hidden Markov tree model to isolate the transient component. We suggest that the auditory system, like the visual system, is probably sensitive to abrupt stimulus changes and that the transient component in speech may be particularly critical to speech perception. To test this suggestion, the transient component isolated by our algorithm was selectively amplified and recombined with the original speech to generate enhanced speech, with energy adjusted to be equal to the energy of the original speech. The intelligibility of the original and enhanced speech was evaluated in eleven human subjects by the modified rhyme protocol. The word recognition rates show that the enhanced speech can provide substantial improvement in speech intelligibility at low SNR levels (8% at -15 dB, 14% at -20 dB, and 18% at -25 dB)

Integer multiwavelet transform has been found attractive in image and video applications where its multiple scaling functions and multiple wavelets provide higher energy compaction in smooth (scaling) subband and smaller values of detail (wavelet) coefficients than using single wavelet. This paper presents a study on an image sharing method applying the integer multiwavelet transform and Shamir's (r,m) threshold scheme that provide highly compact shadows for real time progressive transmission. Experimental results are given to illustrate the characteristics of this method.

Consonants are recognized to dominate higher frequencies of the speech spectrum and to carry more information than vowels, but both demonstrate quasi-steady state (QSS) and transient components, such as vowel to consonant transitions. Fixed filters somewhat separate these effects, but probably not optimally, given diverse words, speakers, and situations, and we suggest that important speech information is contained in transient speech components. To study the relative intelligibility of transient versus steady-state components, we employed an algorithm based on time-frequency analysis to extract QSS energy from the speech signal, leaving a residual signal of predominantly transient components. Psychometric functions were measured for speech recognition of processed and unprocessed monosyllabic words. The transient components were found to account for approximately 2% of the energy of the original speech, yet were nearly equally intelligible. As hypothesized, the QSS components contained much greater energy while providing significantly less intelligibility.

Neurophysical interoperative monitoring (OIM) video is an important tool in biomedical applications and telemedicine and it is essential to maintain high fidelity of the frames even in low bit-rate environment. Toward this end, we studied a spatio-temporal 3D multiwavelet-based OIM video compression system that is capable of high quality output even under the demands of high compression ratios

This paper presents a study of applying an image superresolution method to enhance spatial resolution of MRI heart images from a temporal sequence. The particular method is a wavelet-based projection-onto-convex-set superresolution reconstruction. It makes use of the non-stationary effect in the successive images in the sequence to extract some detail information for image reconstruction at a higher spatial resolution. Our experimental result shows better visual quality than that obtained by using the cubic-spline interpolation.

Motion estimation plays an important role in video compression to remove temporal redundancies between successive frames. It is also useful in aiding detection of moving objects. Full-search block-matching (FSBM) is the most preferred algorithm for motion estimation. Frame-level pipelined FSBM architectures have advantages over block-level pipelined architectures in their simpler control and reduced number of memory accesses. In this paper, a frame-level pipelined FSBM motion estimation array processor for large search range p = qN is presented where q/spl ges/ 1/2 and N/spl times/N is the block size.

This paper presents a wavelet-based method for determining a structural primitive in an image with regular texture. The overcomplete wavelet transform and autocorrelation function of its scaling image are used. An efficient peak-finding method is developed to detect peaks in the autocorrelation function for constructing two displacement vectors from which a parallelogram is formed as the structural primitive of the regular texture. Placement of the primitive in the image gives the structural component of the textured image.

The computed radiography (CR) technique, also known as the storage phosphor imaging technique, has evolved to be a major candidate for large-scale implementation of digital radiography during the past decade. In order to obtain a reasonable spatial resolution, the storage phosphor plate used is generally limited in thickness. This leads to X rays being only partially absorbed by the detector. Useful information may be contained in the X rays transmitted through the detector. Multiple-plate imaging techniques may be used to capture and utilize the X rays more efficiently. In this paper, an image fusion method, based on the Rayleigh principle and the Karhunen-Loeve (K-L) transform, is presented for optimizing the signal-to-noise ratio (SNR) of the fused image on a pixel-by-pixel basis. Because the multiple-plate images contain the same structural information, the signal components of the images are highly correlated with one another. Thus, the K-L transform is applied to decompose each of the multiple-plate images into an eigen image (the estimated signal) and a residual image (the estimated noise). An average representation entropy measure is maximized for selecting the number of eigen components to be included in the signal estimation. An experimental study, using an anthropomorphic chest phantom, is presented to illustrate pixelwise fusion of multiple-plate images. Experimental results show that the SNR of the fused image was improved by 12-48%, depending upon the anatomical regions of interest in the image.

We have developed a computer vision system, including both facial feature extraction and recognition, that automatically discriminates among subtly different facial expressions. Expression classification is based on Facial Action Coding System (FACS) action units (AUs), and discrimination is performed using Hidden Markov Models (HMMs). Three methods are developed to extract facial expression information for automatic recognition. The first method is facial feature point tracking using a coarse-to-fine pyramid method. This method is sensitive to subtle feature motion and is capable of handling large displacements with sub-pixel accuracy. The second method is dense flow tracking together with principal component analysis (PCA) where the entire facial motion information per frame is compressed to a low-dimensional weight vector. The third method is high gradient component (i.e., furrow) analysis in the spatio-temporal domain, which exploits the transient variation associated with the facial expression. Upon extraction of the facial information, non-rigid facial expression is separated from the rigid head motion component, and the face images are automatically aligned and normalized using an affine transformation. This system also provides expression intensity estimation, which has significant effect on the actual meaning of the expression.

A motion estimation algorithm using wavelet approximation as an optical flow model has been developed to estimate accurate dense optical flow from an image sequence. This wavelet motion model is particularly useful in estimating optical flows with large displacement. Traditional pyramid methods which use the coarse-to-fine image pyramid by image burring in estimating optical flow often produce incorrect results when the coarse-level estimates contain large errors that cannot be corrected at the subsequent finer levels. This happens when regions of low texture become flat or certain patterns result in spatial aliasing due to image blurring. Our method, in contrast, uses large-to-small full-resolution regions without blurring images, and simultaneously optimizes the coarser and finer parts of optical flow so that the large and small motion can be estimated correctly. We compare results obtained by using our method with those obtained by using one of the leading optical flow methods, the Szeliski pyramid spline-based method. The experiments include cases of small displacement (less than 4 pixels under 128/spl times/128 image size or equivalent displacement under other image sizes), and those of large displacement (10 pixels). While both methods produce comparable results when the displacements are small, our method outperforms pyramid spline-based method when the displacements are large.

A textured image is considered to be produced by a smooth signal, representing in part the illumination environment, modulated by some frequencies determined by surface characteristics of the textures. Through expansion of the smooth signal into wavelets, a textured image may be decomposed into modulated "wavelets" providing multiresolution information. It is shown that the corresponding wavelet coefficients can be obtained efficiently by using the standard wavelet transform while the associated h and g filters are modulated accordingly. A set of multichannel filters is designed through the use of a modulated "wavelet' multiresolution decomposition, providing both spatial frequency and orientation selectivity. Based on these multichannel amplitude responses as discriminating features, an image containing multiple textures can be effectively segmented. The potential of this approach is shown by experimental results.

The computational issues for symmetric wavelet transforms are investigated. We present a novel frequency-domain algorithm using the discrete cosine transforms (DCTs) and discrete sine transforms (DSTs). A high efficiency is achieved when this algorithm is applied to signals of finite duration, especially images.

A new approach is presented for time-series modeling and prediction using recurrent neural networks (RRNs) and a discrete wavelet transform (DWT). A specific DWT, based on the cubic spline wavelet, produces a set of wavelet coefficients from coarse to fine scale levels. The RNN has its current output fed back to its input nodes, forming a nonlinear autoregressive model for predicting future wavelet coefficients. A predicted trend signal is obtained by constructing the interpolation function from the predicted wavelet coefficients at the coarsest scale level, V/sub 0/. This method has been applied to intracranial pressure data collected from head trauma patients in the intensive care unit. The method has been shown to be more efficient than one which uses raw data to train the RNN.

A dynamic learning rate for back-propagation training of artificial neural networks is proposed as a weighted average of direction cosines of the incremental weight vectors of the current and previous steps. Experiments on training an EEG-based sleep state pattern recognition scheme have demonstrated its improved performance.>

A scale-recursive Kalman filtering on multiscale images from coarse scale to fine scale is proposed which is a generalization of Chou and Willsky's (see Tech. Report CICS-P-333, Center for Intelligent Control System, MIT, May, 1992) model. Through the use of wavelet packet transform, an efficient, highly parallel algorithm has been developed. This scale filtering method can be used to statistically fuse multisensor multiresolution images.>

The paper addresses the idea of using Haar wavelets to generate multiresolution versions of a binary object's surface and to determine the dominant points of that surface. Finding the dominant points is equivalent to segmenting the surface into its piecewise linear components. Two dimensional and three dimensional segmentation algorithms are described and segmentation results are presented for simple 2D and 3D objects. The performance of this approach is compared for the 2D case to two other established methods.>

An edge detection and surface reconstruction algorithm in which the smoothness is controlled spatially over the image space is presented. The values of parameters in the model are adaptively determined by an iterative refinement process; hence, the image-dependent parameters such as the optimum value of the regularization parameter or the filter size are eliminated. The algorithm starts with an oversmoothed regularized solution and iteratively refines the surface around discontinuities by using the structure exhibited in the error signal. The spatial control of smoothness is shown to resolve the conflict between detection and localization criteria of edge detection by smoothing the noise in continuous regions while preserving discontinuities. The performance of the algorithm is quantitatively and qualitatively evaluated on real and synthetic images, and it is compared with those of Marr-Hildreth and Canny edge detectors.>

A multiscale edge detection scheme using a first-order regularization filter (R-filter) is presented. After considering relationships with some of the optimal edge operators, it is shown that the first derivative of the R-filter, which is derived from the 1-D string functional has good edge localization performance. By utilizing this property, the detection performance is improved by combining multiscale edges through a simple yet effective integration scheme. The performance of the algorithm on synthetic and real images is evaluated by using both quantitative and qualitative measures, and compared with that of Canny edge detector.>

An edge detection algorithm based on the regularization theory in which the smoothness is controlled spatially over the image space is presented. The algorithm starts with an oversmoothed regularized solution and iteratively refines the surface around discontinuities using the knowledge on the structure of discontinuities exhibited in the error signal between the image data and the previous regularized solution. The spatial control of smoothness is shown to resolve the conflict between detection and localization criteria. The adaptive nature of the algorithm eliminates the need to select image-dependent parameters and enables the extraction of multiscale features from the image. The computational aspects of the algorithm as well as its performance on real and synthetic images are considered.>

An approach to edge detection problems using regularization theory is presented. The energy functional in the standard regularization has been modified to control the smoothness over the image spatially in order to obtain accurate location of edges. An algorithm which iteratively improves the solution in discontinuous regions by updating the space-varying regularization parameter has been developed. The regularization parameter is controlled by features extracted from the error signal between the regularized solution obtained in the previous iteration and the image data. The algorithm smooths the noisy image without degrading discontinuities. It offers computational advantages and an efficient alternative to existing algorithms for edge detection and for surface reconstruction. The application of the proposed algorithm to various synthetic and real images is discussed.>

A study on applying the discrete pseudo-Wigner distribution to classification of the plasma cortisol short-time series of two disease categories is presented. The autocomponent selection in the Wigner distribution has been studied in regard to feature selection for the pattern recognition. Three energy density features are selected along with the mean level of each time series. These features allow a simple linear classification of the available data. The results show a great potential for using Wigner spectra for feature extraction.>

The discrete smoothed pseudo Wigner distribution is used to process the cortisol time series of both pituitary and ectopic origins for possible visual discrimination in the time-frequency domain. This technique considers the smoothed pseudo Wigner distribution as an image representation. It compares the results of the smoothed pseudo-Wigner distribution after two different filtering processes: one uses a simple averaging filter, and the other uses a preprocessing filter. If their ratio is greater than a selected threshold and the original smoothed pseudo Wigner distribution is positive, then the latter is considered as an autocomponent. Otherwise, it is considered as a cross component and is set to zero.>

The measurement of left ventricular function is greatly enhanced when the three-dimensional shape of the ventricle can be closely assessed by biplane angiocardiograms taken from two perpendicular X-ray camera sets. The ventricle shape can be reconstructed by dividing each of the two orthogonal projection images into corresponding slices, processing each pair of corresponding slices to form a binary cross-sectional image, and stacking up these parallel cross-sectional images. Existing methods of reconstructing cross sections often involve assumptions of elliptic shape or convex symmetry. Recently, a new approach was proposed to reconstruct each cross section by estimating an equal-divisor curve instead of estimating the boundary directly. The approach works well as long as a certain regularity condition (a much relaxed assumption) is satisfied. The notion is extended to the equal-divisor surface of a three-dimensional object with two orthogonal projections. A method for the direct three-dimensional reconstruction of the left ventricle (instead of piling up the parallel reconstructed two-dimensional cross sections) is given.>

A result is presented on information-preserving compression of medical NMR (nuclear magnetic resonance) images for archiving purposes. Both Lynch-Davisson coding and linear predictive coding have been studied. For NMR images of 256*256*12 resolution, the Lynch-Davisson coding with a block size of 64, as applied to prediction error sequences in the Gray code bit planes of each image, gave an average compression ratio of 2.3:1 for fourteen images under test. The predictive coding with a third-order linear predictor and the Huffman encoding of the prediction error gave an average compression ratio of 3.1:1, while the maximum compression ratio achieved was 3.5:1.>

A precision digital edge-line-detection method is presented that was developed for extracting edge contours of resist lines of submicrometer width as imaged by scanning electron microscopy, as a means of inspection in integrated circuit fabrication. The method is based on a modification of the Hough transform.>

The problem of speedily and reliably interpreting range data for industrial object recognition is becoming critically important in the field of robotics and computer vision. This paper presents one approach to attack this problem. It attempts to perform as much classification analysis as possible during the off-line learning process, while only a very small subset of the range data needs to be processed for the on-line recognition. An adaptive matched filter is developed to extract an object's surface feature vectors using range measurement and surface normals. A sparse representation for each object category j is constructed for classification purpose in the form of a Rj-table of selected feature vectors. The approach, based on the concept of the generalized Hough Transform, classifies an object by examining the maximum votes which it receives from various Ri-tables. An optimal selection rule is established for minimizing the misclassification probability. An experiment with a set of simulated range images of nine categories demonstrated the success of the proposed methodology.

This paper presents our progress on automated inspection of VLSI resist patterns as imaged by scanning electron microscopy. A precision digital edge line detection method has been developed for extracting edge contours of resist lines of submicron width, which are crucial for micro registration and detection of defects including bubbles and holes, resist stringers and resist edge defects.