Research On Scale-Invariance Based Methods For Image Feature Extraction

As one of the hot topics in image feature extraction technology, the research on scale invariance based feature extraction methods has gained much attention. Aiming at general image processing tasks, its key idea is to extract scale-invariant key points of one image, and then use these points to create image features which are invariant to shifting and zooming. In light of the theory of image filtering and feature extraction, this thesis mainly deals with the scale-invariant methods for image feature extraction and classification. Firstly, for the research of image filtering, several classic image filtering algorithms are introduced systemically, with their ideas, steps and characteristics being analyzed. Regarding respective limitation of Gaussian filter and Gabor filter, a hybrid filter scheme is presented which combines Gaussian filter and Gabor filter into a general framework. The experiments for image enhancement under different evaluation systems verify that the proposed scheme outperforms conventional filtering methods in terms of information saliency and visual effects, providing theoretical basis for the following improvements on scale-invariant feature extraction methods. Secondly, feature extraction methods based on scale-invariance theory are discussed. Since Scale Invariant Feature Transform(SIFT) tends to produce small, unevenly distributed, and even error features points, a Gabor based SIFT algorithm(GSIFT) is proposed which utilizes above hybrid filter model. By replacing Gaussian filter with hybrid filter when generating the scale space, GSIFT ensures key points extracted are of scale-invariance while the important image information in scale space is well preserved, resulting more rich, uniform and accurate image features. Extensive experiments of image matching on several benchmark image databases verify that the proposed algorithm has better performance in feature extraction, compared with classic SIFT. It also offers high robustness against illumination and affinity. Lastly, the research on image classification based on scale invariant image features is conducted. As the features extracted by GSIFT also contain redundancy, several classic feature dimensionality reduction algorithms are applied to reduce the dimension of GSIFT features, followed by a KNN classifier. The experiments of image classification on several benchmark image databases and measured radar emitter signals verify that this scheme can significantly reduce redundancy of extracted GSIFT features, while core scale invariance of image features are remained. Besides, when combined with dimensionality reduction algorithms, GSIFT still gets better classification results than SIFT related methods.