Research On Image Registration Based On Image Local Invariant Feature Extraction

Image registration is one of the important techniques in pattern recognition and image processing. It is widely applied in computer vision, military, remote sensing and medical image processing, etc. The matching images usually have one or more changes in the image source, size, angle and illumination, at the same time they are also impacted by the complex imaging distortion and noise factors. Therefore, image registration techniques still face severe challenges. Image rotation, size, illumination and fuzzy, have no impact on the local invariant feature extraction technology. And compared to the traditional registration technique based on gray-level information, registration method based on local feature has better matching precision and robustness. In recent years, it has become a hot spot and mainstream research. In the process of multi-source image registration, traditional registration algorithm, which is based on local invariant features, exist redundant interference information and the mismatching rate is high in the similar local feature matching. This paper focuses on the solution of the above problems and mainly improvement researches are as follows:(1) SURF (Speeded-Up Robust Features) algorithm is a better local invariant feature detection method. It is suitable for solving the feature detection problem of the multi-source difference images. The traditional SURF detects feature in the whole image, with the increase of image number and size, the complexity of the algorithm has a geometric ratio growth. In view of the extracted feature points, which have an impact on image matching, are located in the objective scenery, so we put forward improvement ideas, that is, we obtain the edge area of objective by the image edge segmentation technology and mathematical morphology, and then apply SURF method to extract feature in these edge areas. Experimental result shows that SURF method based on the objective edge area is more available. It not only reduces the detection areas and time, but also reduces the interference of the redundant features.(2) The local feature descriptor can describe different characteristics of the specific scene in images, but the feature points, which scattered in the similar local structure, are often mismatched. Inspired by the global feature descriptor can reflect the whole image information, we put forward the idea of combining the global descriptors with local descriptors to describe the space distribution of current feature points. Global distribution descriptors only can be used to describe the feature points which were extracted in the edge areas. We introduce a concept of edge integral image to the global distribution descriptor. And the edge integral image has a low sensitivity to light. When calculate the edge integral image, we only traverse the original image at a time. So the computational complexity is not high. Through the experiments of multiple sets of images, we found that this method does not increase the time complexity, and achieve the purposes of strengthening the uniqueness of the feature points and reducing the mismatching rates.(3) In order to verify the validity of the improved algorithm, we put the improved SURF registration algorithm applied to image stitching, and for splicing traces possibly generate in the process, we use the methods of interpolating technique and multiband image fusion algorithm to splice images, and then adjust the brightness of splicing images. Experimental result shows that our methods have a good effect.