| In this dissertation, the problem of corner-edge feature detection and reconstruction by using virtual line scanning method, as well as the problem of dynamic image/feature classification and identification using a non-iterative neural network, have been studied both theoretically and experimentally, which are organized into 2 parts. In part I, the main objective of solving the corner-edge feature detection problem is to detect accurately the corner points and their connecting lines of any straight-edge objects contained in a digital image, and store the information in a very compact data file. This data file can then be recalled very efficiently to reconstruct the skeleton, or the main features, of the original object in the image. It can be verified by superimposing the reconstructed skeleton onto the original image to determine the accuracy of this corner-edge feature detection scheme.; For the dynamic image/feature classification and identification problem in part II, it is intended to use a non-iterative neural network (NINN) to learn the dynamic variation ranges of some corresponding features in two very similar digital images, and then classify any untrained, time-varying image according to its dynamic variation ranges. The theory of the non-iterative neural network will be reviewed first. Then, the key derivation steps and the major properties of NINN are summarized at the end of Chapter 5. The application of NINN to learn the extreme boundaries and to recognize the untrained patterns varying between these boundaries are studied theoretically and verified experimentally in Chapter 6. |
