Feature Extraction and Signal Processing Approach to Realtime Pattern Recognition.

Abstract

Abstract : The integration of feature extraction for pattern recognition and the digital signal processing into one study as performed in this project has resulted in advances in both areas, and the discovery of many new ideas which are beneficial to both areas. There are common problems, such as the finite sample size effect, in both pattern recognition and signal processing. For example, digital signal processing techniques are much needed in extracting effective features while statistical pattern recognition can be useful in image processing. More specifically, this research has carefully examined the fundamental problem of the finite sample size and its effect on feature selection and classification rules. Most effective features for seismic pattern recognition have been developed through the signal modelling study. In the image recognition work, new results include the rotationally invariant digital Laplacian operation and a new adaptive Kalman filtering technique for efficient realtime image processing. Detailed computer results have been developed and documented to support the theoretical study. Finally for image classification, the specific problem of contextual information is examined and a decision tree procedure is developed which can process both the statistical and structural features for effective classification. (Author)