[论文解读] On Feature based Delaunay Triangulation for Palmprint Recognition

Zanobya N. Khan, Rashid Jalal Qureshi+1 more|arXiv (Cornell University)|Feb 5, 2016

Biometric Identification and Security参考文献 18被引用 8

一句话总结

该论文提出了一种基于特征的Delaunay三角剖分方法用于掌纹识别，通过Niblack二值化提取掌纹线，利用线段端点的Delaunay三角剖分构建拓扑结构，并计算对旋转和缩放不变的相对几何与定量特征。该方法在大规模数据库上使用加权相似性度量和k-NN分类实现了90%的识别率。

ABSTRACT

Authentication of individuals via palmprint based biometric system is becoming very popular due to its reliability as it contains unique and stable features. In this paper, we present a novel approach for palmprint recognition and its representation. To extract the palm lines, local thresholding technique Niblack binarization algorithm is adopted. The endpoints of these lines are determined and a connection is created among them using the Delaunay triangulation thereby generating a distinct topological structure of each palmprint. Next, we extract different geometric as well as quantitative features from the triangles of the Delaunay triangulation that assist in identifying different individuals. To ensure that the proposed approach is invariant to rotation and scaling, features were made relative to topological and geometrical structure of the palmprint. The similarity of the two palmprints is computed using the weighted sum approach and compared with the k-nearest neighbor. The experimental results obtained reflect the effectiveness of the proposed approach to discriminate between different palmprint images and thus achieved a recognition rate of 90% over large databases.

研究动机与目标

• 开发一种基于拓扑与几何特征的鲁棒掌纹识别系统。
• 确保掌纹表示对旋转和缩放的不变性。
• 通过基于特征的三角剖分方法提高掌纹图像之间的区分度。
• 在大规模掌纹数据库上实现高识别准确率。

提出的方法

• 使用Niblack二值化算法对掌纹图像进行预处理，以提取局部特征。
• 检测提取出的掌纹线段的端点，并将其作为Delaunay三角剖分的顶点。
• 通过三角剖分线段端点形成拓扑结构，为每张掌纹生成唯一的网格。
• 从生成的三角形中提取几何与定量特征，包括角度、边长和面积比。
• 相对于拓扑结构对特征进行归一化，以确保对旋转和缩放的不变性。
• 使用特征的加权和计算掌纹对之间的相似性，随后进行k-NN分类。

实验结果

研究问题

• RQ1 Delaunay三角剖分能否有效表示掌纹特征的拓扑结构？
• RQ2 如何使从三角剖分掌纹中提取的几何特征对旋转和缩放保持不变？
• RQ3 所提出的基于特征的方法在多大程度上提升了相对于现有方法的识别准确率？
• RQ4 加权相似性度量能否有效区分不同的掌纹样本？

主要发现

• 该方法在大规模掌纹数据库上实现了90%的识别率。
• 使用相对几何特征确保了对旋转和缩放的不变性，提升了鲁棒性。
• Delaunay三角剖分成功为每张掌纹生成了独特的拓扑结构。
• 加权和相似性度量结合k-NN分类在区分度上优于基线方法。