Research And Application Of Feature Extraction And Feature Registration

During recent years, keypoints extraction and matching, as one of the most important research area in image processing and computer vision, have experienced a rapid development. More and more researchers are shifting their attention to this area and they have applied this technology to the fields of remote sensing, medical image processing, military research etc. But there are still many problems to be resolved in this area. For the example of SIFT descriptor, which is the most widely used local feature descriptor, SIFT descriptor has a very high time complexity and a very bad performance when images have affine transformation. Aiming to resolve these two problems, this paper propose a new method for key point extraction and matching based on the investigation to SIFT descriptor, SURF detector-descriptor scheme and MSER method,and finally this method is applied to the project of image stitching. The main contributions of this paper are:(1) A deep investigation and comparison of SIFT descriptor, SURF detector-descriptor scheme and MSER method. A complete analysis of SIFT descriptor's two defects - high time complexity and low performance with affine transformation. Propose a new method for key point extraction and matching, this method takes the advantage of fast computation of SURF descriptor and good performance for affine transformation of MSER method.(2) A new method to join MSER method and SUFR method together . There is a big difference between MSER method and SURF detector-descriptor scheme. To integrate the advantages of MSER method to SURF descriptor, we propose a new method to transform MESR region to SURF descriptor based on the investigation of SUFR detector-descriptor scheme and MSER method.(3) A total analysis of our new proposed method. We analyze each part of our new method and compare them to the traditional methods. Experiment results have shown that our new proposed method overcome the defects of classic SIFT descriptor. More experiments proved our new method have a very good performance both in time complexity and accommodation for affine transformation.(4) A software for image stitching. We apply our new proposed method to the program of image stitching in different scenes. Resulting images show our method have a very wide usage.