LFMB-3DFB: A Large-scale Finger Multi-Biometric Database and Benchmark for 3D Finger Biometrics

Weili Yang, Zhuoming Chen, Junduan Huang, Linfeng Wang, Wenxiong Kang

Links

• PDF Attachments: 2021’LFMB-3DFB_Yang et al_.pdf
• Zotero Links: Local library

Contributions

1. LFMB-3DFB is the first 3D finger multi-biometric database with multi-view, multi-spectral finger images and 3D finger model with skin and vein textrues.
2. A corresponding multi-view, multi-spectral finger imaging device is designed.
3. LFMB-3DFB contains multi-type finger data. For example, raw 2D images (vein and skin), reconstruction 3D finger model with skin and vein textures.

Some description

• The author called “dorsal side and ventral side” of the finger.
• Finger vein is internal feature and skin is external feature.
• Multi-camera system to surround the finger with blue and near-infrared light sources to irradiate the finger alternately.

Motivation

1. Single modality tends to have unsatisfactory performance and is vulnerabel to be forged.
2. The combination of all the traits on the finger and restore to a 3D representation identifical can tackle above problems.

Challenges

1. How to acquire all the physiological traits from the finger and reconstruct the 3D fingre model.
2. How to extract highly difcriminative identity features from this 3D holographic finger biometric trait.

Imaging System

• The author called it as The multi-view and mulit-spectral 3D finger imaging system.
• Pictures
  • 上图中红色的点是LED，绿色的点是Blue source light.
• 采集流程：
  • 由STM32进行逻辑控制
  • 6个摄像头逐一激活（可以同时采集可见光和NIR）
  • 采集finger skin图像时，除了激活摄像头对侧的blue light不打开，其余的都lit，为了避免过曝同时保证光照均匀。此时所有LED都不开
  • 采集finger vein图像时，仅激活摄像头对侧的LED lit，其余的灯都不开。
• 其他处理
  • 图像内外参校准，内参使用棋盘格[21]和[15]校准外参
  • 手指边界分割：采用Crose to fine的方式。先利用 Canny edge detection, K-means segmentation and multiscale segmentation with guided filter做分割；再利用logical operations and morphological methods对上述三个结果做融合，其结果为粗分割结果。利用这个粗分割的结果作为pseudo-label训练U-net[12]，将训练好的Unet推理结果用作最终分割结果。作者说得到了很好的分割表现。

LFMB-3DFB Information

• Description
  • Totally, 174 volunteers were collected. The index and middle fingers of both hands are captured.
  • Colloect in one session.
  • For each finger, 6 views with an interval of 30 degree were collected. And each view captured 10 times (10 for vein images, 10 for skin images). As a result, each finger has 60 images in each modality (vein, skin).
  • Also, there are 6,950 3D finger mesh models in this database. (The reconstruction process are detailed in Sec 3.)
  • Raw picture resolution is 1280x800 pixels.
• Insight:
  • Contain 3D finger models with vein and skin textures, even if it is reconstructed by themselives.
  • Large scale. (41,700 vein images, 41,700 skin images and 6,950 3D models.)
  • Contain multi-type fingre traits, such as veins, fingerprint, finger kunckle, fingernail。

- **Drawback**:
	- One session captured, which can not ensure the robustness of the trait and the imaging system.
	- There is a big gap between the number of men and women (about 2:1)
	- From the results, maybe the dataset for multi-view recognition task is too easy?

Experiments

作者这里的实验主要是找几个Baseline，然后比较vein (Ve)、skin (Sk)、geometry information (Ge, point cloud without vein and skin texture.), SLF (score-level fusion of 6 views.)在对应的baseline上的表现，即主要看了看不同类型特征的可区分性。

Evaluation Metrics

• Identification: mAP, Rank-1 acc, Rank-5 acc, CMC (Cumulative Characteristic) curve.
• Verification: ROC, EER, TAR@FAR=0.01, TAR@FAR=0.001.

Baseline methods

• Single-view baseline: MobileNetV2[13].
• Multi-view baseline: MVCNN
• 3D End-to-end recognitin baseline: PoinNet and DGCNN
• Score-level fusion (SLF) recognition baseline: score fusion based on SVM.

Results

• Single-view recognition
  • Results (Ca~Cf representas differeant cameras)
  • My obsevation
    • The traits under view Ca (finger ventral) and Cd (finger dorsal) is the most discriminative.
    • Why skin trait is more discriminative than vein trait? Is that means FKP more discriminative? Or is FKP provides more discriminative power?
    • Note that the rank-1 and rank-5 acc do nor reach 1 in single-view identification.
    • Fusion is the best.
• Multi-view Recognitin
  • Results
  • My observation:
    • Fusion is the best.
    • Skin train is better than vein trait.
    • vein and skin trait are both have good performance on Rank-1. Is that means this dataset have no research value on recognition or is too easy? Or we can say the difference may too small?
    • In my opinion, it is a good database for the reseach of 3D finger vein reconstruction.

• 3D End-to-end recognition
  • Results
  • My obsevation:
    • DGCNN is better than PointNet
    • The self-reconstructed 3D finger model is indiscriminative than multi-view.
    • Ge information is discriminative to some degree. The rank 1 and 5 acc are not small.
    • The 3D finger vein reconstruction is need to improve.