| 摘 要: 在计算机视觉中,角点检测是许多视觉任务的核心问题之一。二阶高斯方向导数不仅能有效提取不同方向的图像信息,还具备良好的高斯噪声抑制能力。为进一步提升角点检测与描述性能,提出了一种基于二阶高斯方向导数滤波的新算法GoR2D2用于图像角点检测。提出的算法通过在不同方向上进行二阶高斯方向导数滤波提升网络的角点检测与描述性能。该算法能够自适应选择图像尺度,确保准确检测角点,并在不依赖人工阈值的情况下,精确定位角点。所提算法在图像匹配实验中平均匹配准确率达到了88.2%,实验结果表明,GoR2D2算法显著提升了角点检测与描述的鲁棒性和准确性,特别是在复杂场景和变化条件下,表现出优异的性能。 |
| 关键词: 自适应尺度 二阶高斯方向导数滤波 角点检测 旋转不变性 |
|
中图分类号: TP391
文献标识码: A
|
|
| Corner Detectionand Description Basedon Second Order Gaussian Directional Derivativein Deep Learning Framework |
|
QIU Junjie, LU Jin, WANG Haoyi, ZHANG Xinyue, ZHANG Weichuan
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, Xi’an 710021, China)
231612067@sust.edu.cn; lujin@sust.edu.cn; 2488602498@qq.com; 231612071@sust.edu.cn; zwc2003@163.com
|
| Abstract: Image corner detection is one of the core issues in many computer vision tasks. The second-order Gaussian directional derivative not only effectively extracts image information from different directions, but also has good Gaussian noise suppression ability. To further improve the performance of corner and description, in this work, a new algorithm based on second-order Gaussian directional derivative filtering detection was proposed, named Gaussian optimized Reliable and Repeatable Detector and Descriptor (GoR2D2). The proposed algorithm improves the corner detection and description performance of the network by performing second-order Gaussian directional derivative filtering in different directions. This algorithm can adaptively select the image scale, ensure accurate detection of corners, and accurately locate corners without relying on manual thresholds. The proposed algorithm achieves a Mean Matching Accuracy of 88.2% in image matching experiments, and the experimental results show that the GoR2D2 algorithm significantly improves the robustness and accuracy of corner detection and description, especially in complex scenes and changing conditions, exhibiting excellent performance. |
| Keywords: adaptive scale second-order Gaussian directional derivative filtering corner detection rotational invariance |
