Mesh-Based Multi-Degree Feature Matching Method Under Dense Repetitive Textures

作者：Qi Mu, Kai Du, Zhongxiang Liu, Xin Liang · 年份：2024 · DOI：10.1109/icccs61882.2024.10602871 · 研究领域：Advanced Image and Video Retrieval Techniques、Robotics and Sensor-Based Localization、Image Retrieval and Classification Techniques

In response to the challenge of high similarity among descriptors under dense and repetitive textures, leading to a significant number of false matches during the matching process, this paper presents a grid-based multi-support feature matching method tailored for dense repetitive textures. The method begins by dividing the image into finer grids to reduce the support for false matches. It then expands the support region for matches pending verification and assigns different weights to the support degree of matches based on their distance. Finally, a multi-threshold approach is employed to rigorously filter true matches. To verify the effectiveness of the proposed method, both subjective and objective evaluations were conducted. In subjective evaluations, the method significantly eliminated false matches in scenes with dense repetitive textures. Objectively, the method demonstrated improved matching accuracy on the Boat, Zhantan Temple, Graf, and Leuven sequences by 7.02%, 4.51%, 15.65%, and 6.66% compared to the classical GMS algorithm. Furthermore, the method maintained a low and stable root mean square error and maximum residual across different image sequence scenarios.