Image registration method based on SIFT features

Abstract

The invention discloses an image registration method based on SIFT features. The method includes the following steps: S1. scale-space extremum detection: determining the position and scale of feature points; S2. key point accurate positioning: obtaining accurate positions, scales and ratios of curvatures through conducting data fitting on peripheral data of determined feature point candidate points; S3. main direction resolution: determining designated direction parameter of each key point in accordance to the distribution features in the gradient direction of neighboring pixels of key points; S4. description of generation f vectors: imparting different weights to each sample points with the Gaussian weighting function; S5. feature matching: conducting matching on SIFT features, and determining corresponding relationship of features among images: S6. feature searching: using the BBF searching method to obtain the result of feature matching; S7. the use of improved RANSAC algorithm to remove noise points. According to the invention, the method increases the efficiency of searching, uses the improved RANSAC purifying algorithm to substantially lower the noise proportion of data, increases the reliability of the RANSAC result and data purity, and is conducive to the precision and efficiency of parameter resolution.

Description

technical field [0001] The invention relates to the technical field of commonly used image registration methods, in particular to an image registration method based on SIFT features. Background technique [0002] The SIFT feature has translation, rotation, and scaling invariance, and is also resistant to grayscale transformation and affine transformation. It also has better registration performance when the image quality is relatively poor. Video images are images taken when the camera is moving, so the quality of the image is worse than that of ordinary photos. Camera movement will cause scene blur and ghosting; and the resolution of the image is relatively small, generally 320×240 or 640×480 resolution image. However, in some current image registration methods, after the feature matching metric is determined, the easiest way to search for the nearest neighbor feature is exhaustive search, that is, to search all the feature points in the feature set, find the nearest neigh...

AI Technical Summary

Problems solved by technology

The RANSAC algorithm is an algorithm that calculates the mathematical model parameters of the data and obtains effective sample data based on a set of sample data sets containing abnormal data. The traditional RANSAC algorithm has good performance in removing noise points, but it is not efficient enough.

Images