Higher-order neuralnetworks are a variation of the standard back-propagation neural network, using geometrically motivated nonlinear combinations of scene pixel values as a feature space. The effects of varying featu...
详细信息
ISBN:
(纸本)0819405787
Higher-order neuralnetworks are a variation of the standard back-propagation neural network, using geometrically motivated nonlinear combinations of scene pixel values as a feature space. The effects of varying feature size (in number of pixels), scene size, number of features, summation-over-scene versus maximum-over-scene, and number of hidden layers, are examined.
A model of a neural network system for object recognition in grey-level images that is invariant with respect to position, rotation, and scale is developed. The model is based on the theory of D. Noton and L. Stark an...
详细信息
ISBN:
(纸本)0819405787
A model of a neural network system for object recognition in grey-level images that is invariant with respect to position, rotation, and scale is developed. The model is based on the theory of D. Noton and L. Stark and on the concept of smart sensing. A method for visual image invariant representation is proposed. The method allows transformation of primary features into invariant ones which can be used as input signals for a classical neural network classifier of the high-level structure of the recognizing system.
This paper presents an automatic, optically based image segmentation scheme for locating potential targets in cluttered FLIR images as the front-end image processor of a neural network classifier. The advantage of suc...
详细信息
ISBN:
(纸本)0819405787
This paper presents an automatic, optically based image segmentation scheme for locating potential targets in cluttered FLIR images as the front-end image processor of a neural network classifier. The advantage of such a scheme is speed, i.e., the speed of light. Such a design is critical to achieve real-time segmentation and classification for machine vision applications. The segmentation scheme used was based on texture discrimination and employed orientation specific, bandpass spatial filters (wavelet filters) as its main component. By using the proper choice of aperture pair separation, dilation, and orientation, potential targets in FLIR imagery can be optically segmented using spatial filtering techniques. Segmentation is illustrated for glass template slides, and static and real-time FLIR imagery displayed on an LCTV.
The study of artificialneuralnetworks is an integrated research field that involves the disciplines of applied mathematics, physics, neurobiology, computer science, information, control, parallel processing and VLSI...
The study of artificialneuralnetworks is an integrated research field that involves the disciplines of applied mathematics, physics, neurobiology, computer science, information, control, parallel processing and VLSI. This dissertation deals with a number of topics from a broad spectrum of neural network research in models, algorithms, applications and VLSI architectures. Specifically, this dissertation is aimed at studying neural network algorithms and architectures for pattern classification tasks. The work presented in this dissertation has a wide range of applications including speech recognition, image recognition, and high level knowledge processing. Supervised neuralnetworks, such as the back-propagation network, can be used for classification tasks as the result of approximating an input/output mapping. They are the approximation-based classifiers. The original gradient descent back-propagation learning algorithm exhibits slow convergence speed. Fast algorithms such as the conjugate gradient and quasi-Newton algorithms can be adopted. The main emphasis on neural network classifiers in this dissertation is the competition-based classifiers. The well known linear perceptron and its learning algorithm can deal with linearly separable classification problems. We propose two extensions, the generalized perceptron classifier and the multi-cluster classifier. They can perform more complex pattern classification tasks. We also give the corresponding learning algorithms and prove certain convergence properties. Another powerful classification model is the Hidden Markov Model (HMM), a doubly stochastic automaton that has been applied in speech recognition. We propose the Ring Hidden Markov Model (RHMM) and demonstrate its good performance in a shape recognition application. Due to the rapid advance in VLSI technology, parallel processing, and computer aided design (CAD), application-specific VLSI systems are becoming more and more powerful and feasible. In particula
artificialneuralnetworks are systems composed of interconnected simple computing units known as artificial neurons which simulate some properties of their biological counterparts.They have been developed and studied...
artificialneuralnetworks are systems composed of interconnected simple computing units known as artificial neurons which simulate some properties of their biological counterparts.
They have been developed and studied for understanding how brains function, and for computational purposes.
In order to use a neural network for computation, the network has to be designed in such a way that it performs a useful function. Currently, the most popular method of designing
a network to perform a function is to adjust the parameters of a specified network until the network approximates the input-output behaviour of the function. Although some analytical knowledge about the function is sometimes available or obtainable, it is usually not used. Some neural network paradigms exist where such knowledge is utilized; however, there is no systematical method to do so. The objective of this research is to develop such a method.
A systematic method of neural network design, which we call algebraic derivation methodology, is proposed and developed in this thesis. It is developed with an emphasis on designing neuralnetworks to implement imageprocessing algorithms. A key feature of this methodology is that neurons and neuralnetworks are represented symbolically such that a network can be algebraically derived from a given function and the resulting network can be simplified. By simplification we mean finding an equivalent network (i.e., performing the same function) with fewer layers and fewer neurons. A type of neuralnetworks, which we call LQT networks, are chosen for implementing imageprocessing algorithms.
Theorems for simplifying such networks are developed. Procedures for deriving such networks to realize both single-input and multiple-input functions are given.
To show the merits of the algebraic derivation methodology, LQT networks for implementing
some well-known algorithms in imageprocessing and some other areas are developed by using the above mentioned theorems and proc
This paper will review recent advances in the applications of artificialneural network technology to problems in automatic object recognition. The application of feedforward networks for segmentation, feature extract...
详细信息
A system that can be used as a feature extraction unit in a low-level pattern recognition system is described. It is assumed that such a system acts as a linear mapping between the pattern space and the feature space....
详细信息
ISBN:
(纸本)0819405787
A system that can be used as a feature extraction unit in a low-level pattern recognition system is described. It is assumed that such a system acts as a linear mapping between the pattern space and the feature space. It can therefore be completely described by a number of filter kernels. These filter kernels are usually constructed by the designer of the system. In the approach to filter design described in this paper, the filter kernels are not created manually. Instead, the authors feed the system during the training period with a representative selection of the patterns that they want to recognize. During the training phase, a learning rule (based on a quality function) is used to update the current form of the filter functions. After the training period, there is a filter system that is is optimally adapted to the recognition of this particular set of patterns of interest. In the first part of the paper, some results from work on group theoretical filter design as described. Within this framework, optimal filter functions can be constructed for a large class of pattern recognition problems. These analytical solutions can then be compared with the filter functions learned by our system. The overall structure of the system and several variations of the basic model are described. A quality function is introduced, and a learning filter system is described as an optimization process. This leads to update rules that are significantly different from other, similar, systems investigated previously. Finally, the performance of the system with the help of several examples is demonstrated.
artificialneuralnetworks, inspired by the neural structure of the brain, is a rapidly expanding field of research based on algorithms to solve a wide spectrum of tasks including speech recognition, imageprocessing,...
artificialneuralnetworks, inspired by the neural structure of the brain, is a rapidly expanding field of research based on algorithms to solve a wide spectrum of tasks including speech recognition, imageprocessing, planning, optimisation and other pattern processing tasks. Although a growing number of neural models have been developed to support a variety of applications, neural network programming is still mainly done using conventional languages. This thesis investigates the problems concerned with the programming of neural network models and their portability. The main goal of this thesis is to propose and develop a programming system that can facilitate the implementation of a range of neural network models on a range of hardware. This led to the design and implementation of a programming system called NPS, and a specialised neural network implementation language called NIL. NIL, which forms the neucleus of the programming system NPS, is a low level, machine independent network specification language designed to map a spectrum of neural models onto a range of architectures and thus supporting portability. The neural network programming system NPS provides the user with a system consisting of: A programming language, NIL, to specify network models. A utility, to save partially trained networks for further training. o Libraries of functions and algorithms, to aid the network construction and the execution of standard models. The neural network programming language NIL consists of two major components: A network implementation sub-language, which provides mechanisms for specifying the functions of the nodes and the interconnection topology of the network. A manipulation sub-language, which provides interactive control and modification facilities for use during the training and the recall phase of the network. These sub-languages together produce a low level, machine independent network specification language that can be used to port neural network models. Chapte
Tracking single pixel targets in real time is a difficult task. In a previous paper the authors described a filter that detects possible trajectories in high-noise images. A trajectory image is generated and then used...
详细信息
ISBN:
(纸本)0819405787
Tracking single pixel targets in real time is a difficult task. In a previous paper the authors described a filter that detects possible trajectories in high-noise images. A trajectory image is generated and then used by a Hough transform network to detect lines as well as their end point (the target location at the last frame of the sequence used to generate the image). The Hough transform requires too much processing time when implemented in conventional computers. In this paper, a completely different architecture implementable in analog VLSI form which performs the operation in an extremely short time is proposed. The network uses intrinsic properties of the transform in order to map points in the image plane to corresponding points in the parameter plane. Simple summers are used to `connect' all possible image lines to cells in parameter plane. A peak detecting neural net is then used to detect high count cells in the parameter plane. The technique, developed for straight lines, can easily be extended to second order curves at the cost of increased network complexity. The network has been simulated with simulated target trajectories superimposed on real background images (clouds, etc).
A four-stage object recognition system using (1) Morlet wavelets, (2) vector quantization, (3) hierarchical feature building, and (4) graph matching with dynamical link architecture is presented. The main concept is t...
详细信息
ISBN:
(纸本)0819405787
A four-stage object recognition system using (1) Morlet wavelets, (2) vector quantization, (3) hierarchical feature building, and (4) graph matching with dynamical link architecture is presented. The main concept is that of hierarchically building robust features and topologically associating them with stored feature sets (graphs). The system, being comprised of spatially- local processing stages, is invariant to translation and robust with respect to distortion, partial occlusion, and rotation. Preliminary results from each processing stage are presented. A fusion system, derived from the presented paradigm, is proposed.
暂无评论