Article Preview
Top1. Introduction
In recent years, finger vein, as the popular 2nd generation biometric identification tools, has drawn the attention of more and more scholars (Kumar & Zhou, 2012; Lee, Jung & Kim, 2011). At the same time, finger vein image quality has also made certain research progress (Xie, Zhou, Yang, Lu, & Pan, 2013; Peng, Li, & Niu, 2014; Lee, Khalil-Hani, Bakhteri & Nambiar, 2017). Due to individual differences, changes in the collection environment, and differences in the performance of acquisition equipment, the quality of some collected images is not ideal. In the recognition system, inferior images will seriously affect feature extraction and feature matching. Therefore, in order to filter low-quality images and select high-quality images to input finger vein recognition system, it is necessary to realize accurate and rapid quality assessment after collecting the finger vein images.
According to different purposes of establishing the quality assessment models, existing finger vein image quality assessment schemes can be roughly classified to three categories: (1) This method have fused several quality feature parameters which are manually designed by researchers (Ma, Wang, Fan & Cui, 2012; Yang, Yang, Yin & Xiao, 2013; Qin, Chen & He, 2017); (2) the method is performed on the basis of the number of vein points with vein pattern detection (Nguyen, Park & Shin, 2013; Huang, Kang, Wu, Zhao & Jia, 2016); (3) the method is based on feature representation of deep learning (Qin & Yacoubi, 2015; Qin & Yacoubi, 2017). The first method aims to establish the model which is basically consistent with the evaluation effect of human visual system. This kind of method has to first analyze the factors which may affect the quality of finger veins. After that, the corresponding characteristic parameters are proposed and the features that can characterize the quality of the finger veins are designed manually. It is challenging to realize an effective and robust finger vein quality feature extraction. The second method considers that the quality of the finger vein image mainly relates to whether the satisfactory finger vein feature can be extracted, instead of the judgment result of the human visual system. Therefore, whether or not a large number of clear venous points can be detected becomes an indicator of the quality in such methods. Furthermore, some complicated pre-processing work must be carried out in order to detect the vein points accurately, and the detection process is also comparatively time-consuming in such methods. The third method uses the convolutional neural network to assess the images. For example, the literature (Qin & Yacoubi, 2015) have put forward a hypothesis that the image which was erroneously rejected in the verification scheme are poor images, under such hypothesis, images are automatically labeled the quality. In particular, the labeling is not robust in literature (Qin & Yacoubi, 2015) because it involves only one finger vein image authentication system. That is to say, the model from literature (Qin & Yacoubi, 2015) can effectively distinguish whether the finger vein image can be correctly recognized by this verification system, but not necessarily for other identification systems.