Image Stitching Based On Multiple Homographies

Digital image stitching technology has received a great deal of attention in the field of image processing, which can blend multiple images containing the overlapping area into a picture of the complete scene with high resolution and wide field of vision, and has been widely applied to and many fields, such as remote sensing, medical, virtual reality.Image stitching technology is composed by image registration and image fusion, of which image registration is the foundation of image fusion and is the most critical technology. Feature-based image registration has become a mainstream technology in image registration because of its advantages of small amount of calculation, wide scope of application of and robustness. The classic image registration method takes the projection matrix as the registration model, and uses random sampling consensus algorithm to match feature points and solve the parameters of the image registration model.However, the random sampling consensus algorithm can’t guarantee consistent results in each calculation, and this will cause inconsistent image registration model parameter results, leading to changes in image stitching results. This paper studies the instability of results using random sample consensus algorithm to match feature points and calculate the registration model parameters, and analyzes the causes of this instability. Considering the similarity of the feature point descriptors and the consistency of feature points in the spatial distribution of the location, we propose a new way to matching points to remove the mismatching feature point pairs. Finally, the optimized method is used to complete the calculation of the registration model parameters.On the other hand, the traditional image registration method chosen a projection matrix with8parameters as the registration model, but the projection matrix only fits the case that the image contains only one plane. If the image contains more than one plane, using the projection matrix model as the registration will result in incorrect registration results, thereby exhibit misalignment artifacts and so on. For multi-planar image stitching, we propose a nonlinear image registration mode based on multiple homographies which distinguishes feature points on different planes using Gaussian Mixture Model, calculates homographies of this planes with corresponding feature point pairs and form the final registration model by fusing these homographies together. Finally, this paper completes multi-planar images stitching via image registration model based on multiple homographies.