A Research On The Algorithms Of Feature Extraction Of Image Based On Wavelet Theory

The wavelet transform theory and its applications have been developed furiously in the last decades, especially in the field of image processing. These researches indicate that the wavelet transform coefficients of image in different sub-band include abundance messages of image. And different sub-band represents different messages. This thesis is emphasis on how to use this characteristic to extract feature of image.This thesis introduces the theory of wavelet transform, firstly. Then using wavelet transform in the image processing, the thesis researches the feature extraction and rotation-scale invariant of image, and draws out their arithmetic. As far as the feature extraction, a new method to extract the features based on the wavelet multi-scale transform is put forward. This method uses the energy concentricity of the image's wavelet coefficients to select the parts having concentrated energy, and uses these parts to construct the feature vectors. This kind of feature vectors includes the most edge, texture, luminance and some of structure features. It can be used in many kinds of image's feature extraction without any kind of limit .And it can solve the problem that the single feature extraction must deal with the feature's illegibility. In the experiment, normal white noises having different range are added into the images, and the result approved that the feature vectors are robust to noise. All of these facts indicate that this method is feasible and can be used in the image matching and image searching, etc.The rotation-scale invariant is a difficult and hot discussed question in the wavelet that effect wavelet's improving. The method using in this thesis translates the figure into log-polar, to get partly rotation and scale invariant at a cost of row shift. Firstly, asing the adaptive row shift invariant wavelet packet transform, the thesis abstracts a figure's rotation-scale invariant feature. And this arithmetic of the thesis can resolve the wavelet packet transform sensitivity to the rotation-scale invariant in a way, and then wavelet packet transform can be used in the more expansive field.