Wavelet-based image feature extraction
Joerg Meyer, Zhihe Zhang · 2007
Wavelets are compactly supported functions, which are zero outside a finite interval. A wavelet transform provides both the space and frequency localizations of a signal. Wavelets are created from a single function, called the mother wavelet, by stretching and translation. This property makes wavelets a powerful tool in multiresolution representation of signals. This thesis focuses on the development of new ways of extracting features from images. The feature extraction algorithms designed are based on wavelet multiresolution analysis. The feature extraction process consists of two steps, edge detection and feature labeling. In the first step, an image is transformed into three levels of resolution. The detail coefficients at each level are then filtered using a set of arithmetic, logical, and thresholding operations, and finally combined to reveal the edges. In the second step, a 'snake' moves along the edges searching for high curvature points, intersection points, and edge ends, and label them as the object features. The algorithms are able to extract features from images corrupted by noises, thus making them superior to traditional, non-wavelet-based techniques.