In recent years pattern recognition has evolved to a mature discipline and has been successfully applied to various problems. A fundamental part of an automatic pattern recognition system is classification, where a pattern vector is assigned to one of a finite number of classes. This thesis reports on the development and design of pattern classifier algorithms, with particular emphasis on statistical algorithms which employ discriminant functions. The first part of this research work investigates the use of linear discriminant functions as pattern classifiers. A comparison of some well known methods, including Perceptron, Widrow-Hoff and Ho-Kashyap, is presented. Using generalised linear modelling a new method of training discriminant functions is developed. In this method the linear discriminant function is transformed by a non-linear link function which associates with each pattern vector a measure which is bounded in the range of 0 to 1 according to the class membership of the pattern. In simulations the GLM approach is applied both to synthetic data and to experimental data from a binary pattern matching problem. It is seen that GLM exhibits faster and more reliable convergence than existing linear discriminant approaches. Extensions of this method to Piecewise linear discriminant functions and to polynomial discriminant functions are explored. Application of self-organising methods for efficient generation of polynomial discriminant functions is also investigated. In the second part of the work a review of neural networks is presented, followed by an analysis and formulation of a popular neural network training algorithm, namely Backpropagation (BP). The capabilities and deficiencies of BP and its variations are experimentally evaluated by computer simulations. An alternative formulation based on Empirical Maximum Likelihood (EML) is also proposed. This approach is shown to have a simpler error landscape in comparison to the original BP based on mean square error. Simulations show that the EML approach generally provides faster convergence, involves fewer calculations per iteration than conventional BP, and results in equally good classification performance.

Viewdata systems in commercial use at present offer the facility for transmitting alphanumeric text and graphic displays via the public switched telephone network. An enhancement to the system would be to transmit true video images instead of graphics. Such a system, under development in Britain at present uses Differential Pulse Code Modulation (DPCM) and a transmission rate of 1200 bits/sec. Error protection is achieved by the use of error protection codes, which increases the channel requirement. In this thesis, error detection and correction of DPCM coded video signals without the use of channel error protection is studied. The scheme operates entirely at the receiver by examining the local statistics of the received data to determine the presence of errors. Error correction is then undertaken by interpolation from adjacent correct or previousiy corrected data. DPCM coding of pictures has the inherent disadvantage of a slow build-up of the displayed picture at the receiver and difficulties with image size manipulation. In order to fit the pictorial information into a viewdata page, its size has to be reduced. Unitary transforms, typically the discrete Fourier transform (DFT), the discrete cosine transform (DCT) and the Hadamard transform (HT) enable lowpass filtering and decimation to be carried out in a single operation in the transform domain. Size reductions of different orders are considered and the merits of the DFT, DCT and HT are investigated. With limited channel capacity, it is desirable to remove the redundancy present in the source picture in order to reduce the bit rate. Orthogonal transformation decorrelates the spatial sample distribution and packs most of the image energy in the low order coefficients. This property is exploited in bit-reduction schemes which are adaptive to the local statistics of the different source pictures used. In some cases, bit rates of less than 1.0 bit/pel are achieved with satisfactory received picture quality. Unlike DPCM systems, transform coding has the advantage of being able to display rapidly a picture of low resolution by initial inverse transformation of the low order coefficients only. Picture resolution is then progressively built up as more coefficients are received and decoded. Different sequences of picture update are investigated to find that which achieves the best subjective quality with the fewest possible coefficients transmitted.

While automated storage and retrieval systems for textual and numeric data are now commonplace, the development of analogous systems for pictorial data has lagged behind - not through the lack of need for such systems, but because their development involves a number of significant problems. The aim of this project is to investigate these problems by designing and evaluating an information retrieval system for a specific class of picture, 2-dimensional engineering drawings. This involves consideration of the retrieval capabilities needed by such· a system, what storage structures it would require, how the salient features of each drawing should be represented, how query and stored shapes should be matched, what features were of greatest importance in retrieval, and the interfaces necessary to formulate queries and display results. A form of hierarchical boundary representation has been devised for stored shapes, in which each boundary can be viewed as a series of levels of steadily increasing complexity. A set of rules for boundary and segment ordering ensures that as far as possible, each shape has a unique representation. For each level at which each boundary can be viewed, a set of invariant shape features characterizing that level is extracted and added to the shape representation stored in the database. Two classes of boundary feature have been defmed; global features, characteristic of the boundary as a whole, and local features, corresponding to individual fragments of the boundary. To complete the shape description, position features are also computed and stored, to specify the pattern of inner boundaries within the overall shape. Six different tYPes of shape retrieval have been distinguished, although the prototype system can offer only three of these - exact shape matching, partial shape matching and similarity matching. Complete or incomplete query shapes can be built up at a terminal, and subjected to a feature extraction process similar to that for stored drawings, yielding a query fIle that can be matched against the shape database. A variety of matching techniques is provided, including similarity estimation using global or local features, tests for the existence of specified local features in stored drawings, and cumulative angle vs distance matching between query and stored shape boundaries. Results can be displayed in text or graphical form. The retrieval performance of the system in similarity matching mode has been evaluated by comparing its rankings of shapes retrieved in response to test queries with those obtained by a group of human subjects faced with the same task. Results, expressed as normalized recall and precision, are encouraging, particularly for similarity estimation using either global or local boundary features. While the detailed results are of limited significance until confrrmed with larger test collections, they appear sufficiently promising to warrant the development of a more advanced prototype capable of handling 3-D geometric models. Some design aspects of the system would appear to be applicable to a wider range of pictorial information systems.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.