K.M. Curtis

Shape recognition using a novel fractal technique

Fractal transformations are an exciting new scale invariant shape recognition technique developed by the authors. They have been successfully used for spatially invariant shape recognition using a cumbersome rotational invariance algorithm. Within this paper a new rotational invariance algorithm is presented which makes the recognition process inherently rotationally invariant. A full discussion of fractal recognition is given and previous results discussed. The new algorithm is detailed and applied to a difficult recognition problem. The results of this application show that the new algorithm is successful in achieving the desired effect.

Machine parts classification based on a digital neural network

This paper describes the application of a digital neural network, based on the probabilistic version of the RAM neuron (pRAM), to image processing. The most important pRAM controlling parameters are discussed, along with the application of two types of learning algorithm, based on reinforcement learning and data analysis. The performance of the system is evaluated with respect to its classification of machine parts within a black and white image.

Fuzzy recognition of cricket batting strokes based on sequences of body and bat postures

A mathematical classification of two cricket batting strokes, using fuzzy set theory, is presented. A system is proposed, to capture the motion of a batsman/batswoman whilst playing a stroke. This is then compared to known strokes, provided by the classification and feedback is provided which outlines how well the selected stroke was played. This proposed expert system uses motion sensors in the classification process.

Real time analysis for the synthesis of high quality male, female and child speech

The analysis and synthesis of speech are two problems that have long been investigated. In order that the successful machine production of high quality speech can be realised, these problems have to be addressed jointly if a satisfactory conclusion is to reached. This paper presents an integrated analysis and synthesis system, developed to complement each other, resulting in high quality, multi lingual and dialectal speech synthesis and analysis in real time.

Multiple channel crosstalk removal using limited connectivity neural networks

Limited connectivity neural network architectures are investigated for the removal of crosstalk in systems using mutually overlapping sub-channels for the communication of multiple signals, either analogue or digital. The crosstalk error is modelled such that a fixed proportion of the signals in adjacent channels is added to the main signal. Different types of neural networks, trained using gradient descent algorithms, are tested as to their suitability for reducing the errors caused by a combination of crosstalk and additional gaussian noise. In particular we propose a single layer limited connectivity neural network since it promises to be the most easily implemented in hardware. A variable gain neuron structure is described which can be used for both analogue and digital data.

Dynamically deactivating hidden neurons in a multilayer perceptron neural network

In this paper we present an approach that terminates the processing of hidden nodes, within a multilayer perceptron (MLP) neural network, if they become inactive during the learning process. The determination of the activity and non-activity of hidden nodes are based on the mean deviation of changes in the average derivative of the hidden nodes within an interval of several iterations. A decreasing threshold value is used to evaluate the mean deviation and hence to deactivate the hidden nodes accordingly.

Weight and neuron output resolution requirements in optoelectronic neural networks

We demonstrate that the resolution requirements of the optoelectronic devices used in the communication links of an analog multiperceptron neural network, trained with the standard backpropagation algorithm, can be simultaneously reduced to 8 bits (receiver) and 4 bits (transmitter), respectively, without any significant effect on the networks learning or generalization performances. In addition, we also show that a simple modification to the sigmoidal function, used within each neuron architecture, permits the resolution requirements of the optoelectronic receiver to be further reduced to 4 bits without any additional effect on network performance other than a reduction in learning rate. Both of these limited device resolution performances, however, can be achieved only provided that the weight-storage and the weight-updating procedures are maintained at 14 bits or greater.