2019TPAMI Contactless biometric identification using 3D finger knuckle patterns

Contactless Biometric Identification using 3D Finger Knuckle Patterns

1. Introduction

1. 一般，关节模式的最具有判别性的信息是：knuckle curves and creases。但是从2D的图像准确的提取出这些特征信息，比较困难，因为光照的改变对其影响很大.
2. 3D的指关节信息更可靠，由于3D信息对于光照的改变是没有变化的。
3. 可以同时利用2D和3D的信息，来提高性能
4. challenges：
   1. it’s difficult to design a feature descriptor to robustly recover unique information from 3D curve and creases.
   2. 3D 扫描技术的限制，相对于方便的2D系统，3D技更加 耗时耗力。
   3. 瓶颈：缺少数据库

2. Related Work

1. Based on the nature of feature descriptors, these approaches can be largely categorized into three categories:
   1. those based on subspace learning
   2. spectral features
   3. those based on the discretization of local features

Key contributions

1. Simultaneous acquisition of 3D and 2D finger knuckle images
2. We develop a new feature descriptor to efficiently and more accurately match 3D finger knuckle biometric patterns
3. this paper develops the first two-session 3D finger knuckle database

3. 3D finger knuckle identification

3.1 Image acquisition

1. 我们通过光度立体视觉方法（photometric stereo approach）方法获得3D手指关节图像。
2. 一个low-cost 固定的摄像机+7个围绕在摄像机分布的灯+控制电路
3. 灯的位置可以通过，放置一个大头针，根据阴影的高度等来近似确定。

3.2 Image Preprocessing and segmentation

1. 提取感兴趣的区域，即包含手指关节图像的区域。

   用一个固定大小的矩形，沿着长宽方向进行寻找，最终，确定矩形像素最多的区域作为分割区域

2. 分割后的区域进行对比度拉伸。

3.3 3D Reconstruction