This paper presents a 10-bit two-channel time-interleaved successive-approximation-register (SAR) analog-to-digital converter (ADC) to achieve a maximum sampling rate of 400 MHz in 90-nm CMOS. The channel-ADC incorporates a 2b/cycle-assisted SAR architecture to effectively speed up its operation speed. A new dynamic register is proposed to reduce the DAC control delay. To maintain small capacitor arrays, a multiple-reference C-DAC is applied to avoid using a tiny unit capacitance. This prototype ADC consumes a total power of 3 mW from a 1.2-V supply at 400-MS/s. Measured DNL and INL are -0.77/+1.2 LSB and -1/+1 LSB, respectively. The measured peak SNDR and SFDR are 51 dB and 62.3 dB, respectively. The measured ENOB is 8.2 bits, equivalent to a figure-of-merit of 25.5 fJ/conversion-step.

This paper proposes a new neural network model, a multi-resolution fuzzy ART (MRF-ART), which employs fast competitive learning and efficient parallel matching to solve complex data classification problems. The architecture of MRF-ART not only preserves the ART-type neural network characteristics but also extends their capability to represent input patterns in a hierarchical fashion. To achieve this, an MRF-ART network uses multiple output layers arranged in a cascaded manner which is completely different from a conventional fuzzy ART network with only one output layer. Moreover, the parallel matching process enables the parallel hardware implementation of an MRF-ART. To demonstrate the data representational capability of an MRF-ART network, we applied it to two data sets and the results indicated that fine-to-coarse data representation can be achieved.

An automated method is proposed for anatomic standardization that elastically map one subject's MR image to a standard reference MR image to enable inter-subject and cross-group studies. In this method, linear transformations based on bicommissural stereotaxy are first applied to grossly align the input image to the reference image. Then, an active contour model is employed to refine the contour and surface description of the input image. Based on the contour and surface correspondence established in these previous steps, a non-linear transformation is determined using the proposed weighted local reference coordinate systems to warp the input image. In this method, geometric correspondence established based on contour matching is used to control the warping and the actual image values corresponding to registered coordinates need not be similar. We tested this algorithm on various synthetic and real images for inter-subject registration of MR images.

The estimation of the fractal dimension is essential in fractal-based image segmentation, classification and shape analysis. The most popular estimation approach is based on box-counting. However, the partition and counting methods used in the regular box-counting scheme produces inaccurate results. In this paper, a more accurate fractional box-counting approach is proposed to estimate the fractal dimension in an image. By separating the concepts of base scale and counting scale, the new approach is able to use a deformable box to capture fractal property at some predetermined resolution. Preliminary results using the fractional box-counting approach have demonstrated that it is an accurate and efficient method.

This paper is concerned with the problems of reconstructing a closed surface from scattered, noisy 3-D data. A progressive shell model is a 3-D extension of the 2-D progressive contour model. We employ finite element methods (FEMs) to reduce the number of the required variables and improve the efficiency in storage and computation. The fundamental forms in differential geometry are used to measure rigid-motion invariant properties and formulate the internal energy of the shell. We also develop a wireframe model associated with a subdivision scheme to overcome the difficulty of generating a smooth boundary between two adjacent patches. This is a direct application of the 2-D contour model where curve segments or wires are used instead of patches. In the subdivision scheme, we impose the co-plane constraints to determine a unique normal vector at the interpolated mid-point. To demonstrate the descriptive ability of a wireframe model, we conduct experiments on 3-D data set of a tumor and a face.

This paper presents a progressive contour model based on the idea of deforming a contour from an initial shape as a source of prior knowledge by minimizing a defined contour energy to extract a desired contour from an image. The internal component of the contour energy imposes smoothness constraints on the displacements of deformations and the external component establishes the correspondence between the contour and image edges. A sequence of deformations is determined by repeatedly deforming and updating the initial contour. We show that the contour deformed by this sequence will smoothly and progressively approach a uniquely defined contour. We implement this model using a finite-element method, a multigrid algorithm, and an unconstrained optimization method. This approach offers several attractive advantages including a satisfactory convergence rate and the adaptation of the smoothness constraints.

An interpolation method, morphological field morphing, using contours of organs as the control parameters is proposed to recover the intensity information in the physical gaps of serial cross-sectional images. Contour information derived from a contour-model-based segmentation process is processed and used as the control parameters to warp the corresponding regions in both input images into comparable shapes. In this way, the reliability of establishing the correspondence among different segments of the same organs is improved and the intensity information for the interpolated intermediate slices can be derived more faithfully. In comparison with the existing intensity interpolation algorithms that only search for corresponding points based on intensity information, this method provides more meaningful correspondence relationships by warping corresponding regions in images into similar shapes before resampling to account for significant shape differences.

Proposes an approach applying artificial neural net techniques to 3D rigid motion analysis based on sequential multiple time frames. The approach consists of two phases: (1) matching between every two consecutive frames and (2) estimating motion parameters based on the correspondences established. Phase 1 specifies the matching constraints to ensure a stable and coherent feature correspondence establishment between two sequential time frames and configures a 2D Hopfield neural net to enforce these constraints. Phase 2 constructs a 3-layer net to estimate parameters through supervised learning. The method performs motion analysis based on sequential multiple time frames. It represents an effective way to achieve optimal matching between two frames using neural net techniques. The energy function of the Hopfield net is designed to reflect the matching constraints and the minimization of this function leads to the optimal feature correspondence establishment. The approach introduces the learning concept to motion estimation. The structure of the net provides the flexibility in estimating motion parameters based on information from multiple frames.

For pt. I see ibid., p. 1386-93 (1995). An approach applying artificial neural net techniques to 3D nonrigid motion analysis is proposed. The 3D nonrigid motion of the left ventricle of a human heart is examined using biplanar cineangiography data, consisting of 3D coordinates of 30 coronary artery bifurcation points of the left ventricle and the correspondences of these points taken over 10 time instants during the heart cardiac cycle. The motion is decomposed into global rigid motion and a set of local nonrigid deformations which are coupled with the global motion. The global rigid motion can be estimated precisely as a translation vecto and a rotation matrix. Local nonrigid deformation estimation is discussed. A set of neural nets similar in structure and dynamics but different in physical size is proposed to tackle the problem of nonrigidity. These neural networks are interconnected through feedbacks. The activation function of the output layer is selected so that a feedback is involved in the output updating. The constraints are specified to ensure stable and globally consistent estimation. The objective is to find the optimal deformation matrices that satisfy the constraints for all coronary artery bifurcation points of the left ventricle. The proposed neural networks differ from other existing neural network models in their unique structure and dynamics.

A fuzzy neural network based approach to 3-D heart motion understanding is proposed. Knowledge from cardiologists is used to specify different classes of motion and classification rules. The objective of this approach is to find the decisions for all possible classes of motion in the form of possibilities. Experiments on real data have been conducted to corroborate the proposed techniques.

Describes the recognition subsystem of a computer vision system based on constructive solid geometry (CSG) representation scheme. Instead of using the conventional CSG trees to represent objects, the proposed system uses an equivalent representation scheme-precedence graphs-for object representation. Each node in the graph represents a primitive volume and each are between two nodes represents the relation between them. Object recognition is achieved by matching the scene precedence graph to the model precedence graph. A constraint satisfaction network is proposed to implement the matching process. The energy function associated with the network is used to enforce the matching constraints including match validity, primitive similarity, precedence graph preservation, and geometric structure preservation. The energy level is at its minimum only when the optimal match is reached. Experimental results on several range images are presented to demonstrate the proposed approach.>

The discrete Gabor (1946) transform algorithm is introduced that provides an efficient method of calculating the complete set of discrete Gabor coefficients of a finite-duration discrete signal from finite summations and to reconstruct the original signal exactly from the computed expansion coefficients. The similarity of the formulas between the discrete Gabor transform and the discrete Fourier transform enables one to employ the FFT algorithms in the computation. The discrete 1-D Gabor transform algorithm can be extended to 2-D as well.

An algorithm for extracting and segmenting surface descriptions from stereo images is presented. Compared with the advance of range image segmentation, the progress of surface segmentation from stereo images is slow. The algorithm integrates the boundary-based and region-based approaches to achieve more satisfactory segmentation results. The algorithm yields such a rich description of a scene in terms of global surface patches and closed surface boundaries. These shape primitives are useful for subsequent high-level processing.>

The authors propose knowledge representation and evidence propagation schemes based on multivariate belief functions and present a medical image recognition system to demonstrate the effectiveness of their application to spatial reasoning. The proposed system, which is based on the blackboard architecture, can mimic the reasoning process of a human expert in identifying the anatomical structures in a set of correlated images acquired from X-ray CT and PD- and T/sub 2/-weighted MRI. In the blackboard-oriented system, different kinds of evidence provided by various knowledge sources form a hierarchy of evidential space to which the Dempster-Shafer theory is applied. The multivariate belief functions are used to represent domain specific knowledge such as rules or facts.>

Addresses the problem of segmenting objects into parts using stereo images. There are three components in the part segmentation process: surface segmentation, region grouping, and volumetric models recovery/segmentation. The surface segmentation process segments the image into a set of regions such that each region represents a surface smooth in depth. The region grouping process merges the segmented regions into meaningful parts. Finally the process of volumetric model recovery/segmentation phase recovers the part model and segments that part into smaller parts if necessary. Since multi-shape models have been used to drive the part segmentation process, one can capture more geometric properties of the object and the application domain of the approach is broader than that of the previous approaches.>

The recognition subsystem of a vision system based on a constructive solid geometry (CSG) representation scheme is described. From the image, a description about the scene-precedence graph-is inferred. Object recognition is formulated as matching a scene precedence graph to a model precedence graph which is executed by a constraint satisfaction network.>

A medical image recognition system is described whose reasoning module uses the features of the Dempster-Shafer (D-S) theory such as compatible frames and multivariate belief functions. The proposed expert system, which is based on the blackboard architecture, is capable of mimicking the reasoning process of a human expert in dividing a set of correlated X-ray CT and T-1 and T2-weighted MR images into semantically meaningful entities. In the blackboard-oriented system, different kinds of evidence provided by various knowledge sources form a hierarchy of evidential space to which D-S theory is applied. Several experimental results are given to illustrate the performance of the proposed system.>

A hierarchical approach is proposed for solving the surface and vertex correspondence problems in multiple-view-based 3D object-recognition systems. The proposed scheme is a coarse-to-fine search process, and a Hopfield network is used at each stage. Compared with conventional object-matching schemes, the proposed technique provides a more general and compact formulation of the problem and a solution more suitable for parallel implementation. At the coarse search stage, the surface matching scores between the input image and each object model in the database are computed through a Hopfield network and are used to select the candidates for further consideration. At the fine search stage, the object models selected from the previous stage are fed into another Hopfield network for vertex matching. The object model that has the best surface and vertex correspondences with the input image is finally singled out as the best matched model. Experimental results are reported using both synthetic and real range images to corroborate the proposed theory.

The authors propose a three-layered neural network to perform Huffman-Clowes scene labeling. The proposed neural network uses the topology and the interconnections of neurons to achieve global consistency through propagating local constraints. The problem-solving knowledge is embedded in the topology as well as the connections between neurons in the network. A brief review of the Huffman-Clowes scene labeling scheme is presented. The proposed constraint propagation neural network is described. Several examples are given to illustrate the operation of the network. Time complexity analysis of the network is discussed. A comparison with conventional algorithms is given. The characteristics of the proposed neural networks are discussed.>

The authors propose a hierarchical approach to solving the surface and the vertex correspondence problems in multiple-view based 3-D object recognition systems. The proposed scheme is a coarse-to-fine search process, and a Hopfield network is employed at each stage. Compared with the conventional object matching schemes, the proposed technique provides a more general and compact formulation of the problem and a solution more suitable for parallel implementation. At the coarse search stage, the surface matching rates between the input image and each object model in the database are computed through a Hopfield network and used to select the candidates for further consideration. At the fine search stage, the object models selected from the previous stage are fed into another Hopfield network for vertex matching. The object model that has the best surface and vertex correspondences with the input image is finally singled out as the best matched model. Results of experiments using both line drawings and real range images to corroborate the proposed theory are also reported.>

A CSG (constructive solid geometry) -based system to recognize 3D objects from range images is proposed. The system starts by segmenting an object in a range image into primitive surface patches with the simplest geometric characteristics. The primitive patches are classified as surfaces of primitive volumes such as sphere, cylinders, cubes, or cones, and the order and types of set operations on these primitives are determined. This is achieved by constructing a precedence graph describing the priorities of operations. The recognition is based on matching of precedence graphs. The advantages and disadvantages of the approach are discussed.>

A linking algorithm to fill gaps between edge segments is presented. The gap filling operation is performed in an iterative manner rather than a single step. During each iteration, the noises are also removed gradually. The proposed algorithm starts with connecting the tip ends of two (line or curve) segments with a line segment and then tries to modify the resulting segment by straight-line fitting. The two major phases of the algorithm are detailed step by step. The advantages of using the algorithm are also discussed.>

A syntactic approach to three-dimensional object recognition from a single viewpoint is proposed. The system consists of two major parts: analysis and recognition. The analysis part consists of selecting primitive surface patches and modeling grammar construction. The recognition part consists of preprocessing, image segmentation, visible primitive surfaces identification, camera model estimation, and structural analysis. Attention is given primarily to the analysis and the structural analyzer of this system.

VLSI technology has recently received increasing attention due to its high performance and high reliability. Designing a VLSI structure systematically for a given task becomes a very important problem to many computer engineers. In this paper, we present a method to transform a recursive computation task into a VLSI structure systematically. The main advantages of this approach are its simplicity and completeness. Several examples, such as vector inner product, matrix multiplication, convolution, comparison operations in relational database and fast Fourier transformation (FFT), are given to demonstrate the transformation procedure. Finally, we apply the proposed method to hierarchical scene matching. Scene matching refers to the process of locating or matching a region of an image with a corresponding region of another view of the same image taken from a different viewing angle or at a different time. We first present a constant threshold estimation for hierarchical scene matching. The VLSI implementation of the hierarchical scene matching is then described in detail.

A 3-D object representation scheme which uses surfaces as primitives and grammatical production rules as structural relationship descriptors is proposed. Possible selections of surface primitives are discussed. Examples are given to illustrate the object description method.

The feasibility of applying image processing techniques to metal surface inspection is demonstrated. Two methods for metal surface inspection are described. In the first method, the metal surface reflective power and the metal surface normal are related by a random surface scattering model. The metal surface profile can then be computed from the metal surface normal. The second method applies pattern recognition techniques to classify metal surfaces into classes of different roughness. Methods of feature extraction and classification have been tested experimentally and the performances of different types of classifier have been compared. A two-level tree classifier using nonparametric linear classifiers at each node gives better than 90% correct classification on the testing set.