| 摘 要: 传统的兴趣点检测方法在实际应用中常常会产生大量响应较弱、稳定性不足的关键点,这些弱关键点容易受噪声、视角变化及光照变化的影响,导致后续匹配和重建任务的性能下降。为此,提出了深度学习框架下的双网络一致性筛选方法用于兴趣点检测。该方法结合了两个独立的兴趣点检测网络,利用双检测器之间的共识机制,仅保留同时被两个网络检测到的关键点,以此抑制弱关键点的影响。实验结果表明,所提检测框架在 HPatches数据集上的平均匹配精度(MMA)为88%,整体性能优于对比的主流算法。 |
| 关键词: 兴趣点检测 双网络一致性 图像匹配 关键点筛选 弱关键点抑制 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金项目资助(61801281 |
|
| A Dua-l Network Consistency Filter Method Under a Deep Learning Framework for Interest Point Detection |
|
LIU Kang, LU Jin
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, Xi’an 710021, China)
231612056@sust.edu.cn; lujin@sust.edu.cn
|
| Abstract: Traditional interest point detection methods are often plagued by the generation of numerous keypoints with weak responses and insufficient stability in practical applications. These weak keypoints are susceptible to noise,viewpoint changes, and illumination variations, leading to degraded performance in subsequent matching and reconstruction tasks. To address this issue, A Dua-l Network Consistency Filter Method under a Deep Learning Framework for Interest is proposed for image interest point detection. This method integrates two independent interest point detection networks and leverages a consensus mechanism between the dual detectors, retaining only the keypoints detected simultaneously by both networks to mitigate the impact of weak keypoints. Experimental results demonstrate that the proposed detection framework achieves an Mean Matching Accuracy(MMA) of 88% on the HPatches dataset,outperforming the compared mainstream algorithms. |
| Keywords: interest point detection dua-l network consistency image matching keypoint screening weak
keypoint suppression |
