Study On Feature Point Detection And Its Application To Image Matching

The art of image matching is a critical content in computer vision and digital image processing, which is widely used in domains of object recognition and analysis, robots vision navigation, pattern recognition, medical image analysis, and so on. However, geometry deformation and grayscale distortion between remote sensing images caused by differences of imaging conditions are big problems in image matching. In order to get perfect matching results in many applications, numerous researchers are engaged in the study of image matching technique. In this thesis, the investigation and discuss are carried out on object matching of remote sensing images, especially on feature point based image matching algorithms. Most work is focused on feature point detecting, matching based on position relationships of feature points, and matching based on descriptors of feature points. The main work and innovations in the thesis are listed as follows.1. Significance and necessity of feature point detecting and image matching are analyzed. Detailed conclusions and summaries are given on their classifications, current research states as well as application states, and problems of image matching are analyzed.2. Emphasis research is given on feature point detectors which are based on gray information of the image directly. Analysis is made on several point detectors frequently used in photogrammetry and a hot detector of Sale Invariant Feature Transform (SIFT) detector in computer vision. A series of experimental comparisons and performance evaluations of these detectors are done under the criteria of velocity, accuracy, as well as adaptability.3. Based on the investigation of several existing Hausdorff distances and the genetic algorithm, an improved Hausdorff distance (IHD) and an efficacious genetic strategy of image matching are presented. Then, a matching method based on position relationships of feature points is designed, which takes feature points instead of the matched image, IHD as the matching measure, and genetic algorithm as the matching strategy. Experimental results prove that this method can get correct matching results effectively when there are obvious intensity variation and partial image distortion in remote sensing images.4. Based upon the deep research on SIFT descriptor, a new SIFT descriptor creation method is proposed, which takes advantage of a circle region around the feature point to creat the descriptor, and then a matching method based on the improved SIFT descriptor is designed. Experimental results show that, when there are complex geometry distortions and illuminance variations between visible spectral or radar remote sensing images with similar resolutions, the matching method proposed by this thesis can still obtain statisfied object matching results.