Research on Correction Method of Local Feature Descriptor Mismatch

Abstract

In this paper, the local feature description algorithm increases the number of mismatches in the case of illumination mutation, and the image deformation often does not appear in a single way. Therefore, there is a high requirement for the stability of the image matching algorithm. If a similar structure appears in the local region of the image, the mismatch rate is higher, especially when the feature vector has no feature semantics or position information. The one-to-many, many-to-many mismatched pair has a higher frequency. In general, most of the methods used to circumvent such errors are commonly used to adjust the distance threshold, but the distance threshold is not targeted and the correct matching point pair may be eliminated. Therefore, Based on the KNN and RANSAC methods, this paper further uses geometric constraints to eliminate mismatches. This method can greatly improve the matching accuracy, so that the image descriptor can maintain high accuracy under the local light mutation environment and even the local similar content. Experiments have shown that our method can achieve very high matching accuracy.