ArcVein-Arccosine Center Loss for Finger Vein Verification

Borui Hou, Ruqiang Yan

Links

• PDF Attachments: 2021’ArcVein-Arccosine Center Loss for Finger Vein Verification_Hou,Yan_.pdf
• Zotero Links: Local library

My Comments and Inspiration

This work seems like a combination of ECA-Net+ResNet as backbone. The most “novel” idea is replacing the Euclidean distance in center loss with cosine distance. The motivation and inspiration are not described clearly.

In addition, I question the authenticity of the results. The whole paper is in a bad organization. It is hard to follow, where the spelling errors, capitalization errors are very much.

Contributions and Important Conclusions

The author proposed a new loss function, which replaces the Euclidean distance in center loss with cosine distance.

The author claimed that this Arccenter is are easy to convergency and stable during traning.

Motivation

Not clear. The author only want to improve the performance.

Motivated by center loss and cosine distance (…Orz)

Methods

Introduce from two parts: feature extraction (network) and loss function

Feature extraction

The combination of ECA-Net and ResNet (Res block)

Input: Input images is with $224\times 224\times 3$. They expand the one-channel gray images to three-channels by copying.

Architecture:

Defination of basicBlcok

There are five basicBlocks with a Average pooling at the end.

Loss Function

A combination of Softmax Loss $L_s$ and Arccenter loss $L_{a}c$, formulated as

As experiment IV-C, the paremeter $\lambda$ set as: USM = 0.5, POLY and SDU = 1, self-collected database = 2.

Arccenter: Just replace the Euclidean distance with cosine distance, i.e., original center loss:

$$L_c=\frac{1}{2}\sum _{i=1}^N\Vert f_i-c_{l_i}{\Vert }^2$$

Replace the Euclidean distance with cosine distance, we can obtain

$$L_{ac}=\sum _{i=1}^N\text{arccos}({\theta }_i)$$

where

$$cos({\theta }_i)=\frac{f_i^{\top }{c}_{l_i}}{\Vert f_i\Vert \Vert c_{l_i}\Vert }$$

Experiments

Database information

4 database are used.

SDU and self-collected database: randomly selected half images for training, the other for testing.

USM and POLY: session 1 for training and session 2 for testing.

ROI extraction is based on this paper.

P. Gupta and P. Gupta, “An accurate finger vein based verification system,” Digit. Signal Process., vol. 38, pp. 43–52, Mar. 2015.

Data Augmentation

Online data augmentation, see following table.

Implement details

• Optimizer: stochastic gradient descent algorithm with Nesterov momentum 0.9

• LR: 0.01, divided by 10 after 30 epochs.

• L2 wright deacy: $5\times 10^{-4}$

• bs: 64

• epochs: 100

• Two-fold cross-validation.

Protocol

• Closed-set protocol:

  • See Database information

• Open-set Protocol:

  • For all databases, one-half of classes are selected for traning (maybe all samples are used), the other classes for testing.