Local feature description method, system and device for rotation invariant multi-source images

Abstract

The invention relates to a local feature description method, system and device for rotation invariant multi-source images. According to the method, a descriptor for local feature matching of the rotation invariant multi-source images is constructed based on a multi-scale multi-direction log Gabor filter; and the descriptor not only can adapt to difference between the multi-source images caused bynonlinear brightness change but also has a certain suppression effect on noises, and can adapt to geometric changes between the images, such as translation, rotation and the like, so that the descriptor has stronger robustness, and the descriptor can be more accurate and efficient for the local feature description method supporting rotation invariance.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method, system and device for describing local features of a rotation-invariant multi-source image. Background technique [0002] With the continuous advancement of computer technology and sensor technology, multiple types of sensors can be used to obtain image information for the same scene. We call this type of image multi-source image. In order to make full use of multi-source image information, it is necessary to correct the multi-source images into a unified coordinate system through registration and fusion, and image matching is a key step in registration, which directly affects the quality of registration and fusion. Due to the differences in the working principles and imaging characteristics of various sensors, there may be different degrees of radiation differences and geometric changes in the same scene in multi-source images, which brings great challenges to...

AI Technical Summary

Problems solved by technology

Region-based matching methods, including cross-correlation, mutual information, and HOPC, can achieve high matching accuracy, but this method is only effective for translational transformations, and the matching effect for geometric transformations other than translational transformations is not good.

In order to adapt to image discontinuity, shadows, and occlusions, it is generally necessary to use a larger template to ensure the accuracy of matching, and the time consumption of matching is also relatively large.

Feature-based matching methods, in addition to feature point detection, also need to perform local feature description and similarity measurement to achieve matching between feature points, mainly by constructing local invariant feature descriptors for matching, most feature extraction algorithms are It is based on image gradient information for feature description, such as SIFT, SURF, HOG, and ORB descriptors, but these descriptors are sensitive to nonlinear brightness differences between images, and there are certain limitations.

When using edge features (EOH), contour features (Shape Context), and structural information (LSS) for local feature description, although the problem of nonlinear brightness is solved to a certain extent, it is greatly affected by noise and is usually only applicable to images. The case where the difference between is small, and does not have rotation invariance

[0004] Among the existing local feature description methods that support rotation invariance, SIFT, SURF, etc. mainly obtain the main direction by calculating the main gradient direction, and realize the rotation invariance according to the main direction, but they are not suitable for multi-sources with nonlinear brightness differences. image, and this method has a certain angle error

PIIFD obtains the main direction through gradient inversion, but this method still has angle errors, resulting in a decrease in matching accuracy

However, LIOP does not need the main direction in the calculation process, but sorts the gray value and performs group grouping to obtain the spatial structure information and realize the description of rotation invariance. However, this method is only suitable for gray images with linear brightness changes, not suitable for Images with nonlinear brightness changes, and the method is sensitive to noise

Images