**4. Discussion**

At first, we discuss the image quality of our datasets in comparison with other publicly available finger- and hand-vein datasets. The evaluation results are listed in Table 1. Considering finger veins, our dataset achieved the best results for GCF and HSNR while it is ranked third for Wang17. These results confirm decent image quality in terms of contrast and also a good image quality in terms of vein specific properties. Considering hand veins our 850 nm data set achieved the best results for Wang17 and the second best ones for HSNR, while the 950 nm dataset achieved the best results for HSNR and second best for Wang17. In terms of GCF, both hand vein data sets are only ranked second and third to last, indicating that the general image contrast is lower than for other datasets. However, the vein specific quality metrics still indicate good image quality, despite the inferior image contrast compared to the other datasets. There are several works in the literature [67–70] that confirm that quality metric results do not necessarily have to correlate with recognition accuracy. The recognition accuracy is the most important aspect of a sensor design and dataset, thus we decided to focus on the recognition accuracy evaluation instead of evaluating the image quality only. Also note that other data set and sensor papers do not report the image quality, thus a direct comparison is not possible based on the image quality.

In the following, we compare our recognition performance evaluation results to other results reported in the literature. Matsuda et al. [12] reported an EER of 0.19% for their walk through style finger vein recognition system. Raghavendra et al. [11] et al. reported an EER of 1.74% for their systems in case of finger veins only. Kim et al. [10] arrived at an EER of 3.6% for their NIR laser based contactless acquisition set-up. Sierro et al. [9] did not present any performance evaluation of their dataset. None of the mentioned authors disclosed their dataset, so their results are not reproducible. As we aim for reproducibility, all the results listed in Tables 4 and 5 are evaluated on publicly available finger and hand vein datasets and we base our discussion on those results only.

Table 4 lists performance results (in terms of the EER unless indicated otherwise) achieved on various finger vein datasets, ordered by the year of publication, where the last row corresponds to the results presented in this work. The listed results are the best reported ones from the original dataset authors given that they were available and indicated by "-" if they were not available. Note that all of the listed datasets, except the one presented in this work, have been acquired in a non contactless way. Especially compared to the PLUSVein-FV3 dataset, which has been acquired using the same type of NIR enhanced camera and NIR laser modules, the results on our proposed dataset are clearly inferior (5.61% EER vs. 0.06% EER). However, given the increased level of difficulty and challenges of this new, contactless finger vein dataset, the results are still in an acceptable range. The proposed acquisition device design and thus, the acquired dataset, allows for more degrees of

freedom in terms of finger/hand movement and thus, more unrestricted finger/hand positioning. This introduces different kinds of finger/hand misplacements, including tilts, bending, in-planar and non-planar rotations. Prommegger et al. showed, that especially those kinds of misplacements cause severe performance degradations for other publicly available datasets [7,71], especially for the SDUMLA-HMT finger vein dataset [60]. Also note that our results should only serve as a baseline and allow room for further improvements. We did not apply any special kind of finger misplacement corrections. If advanced correction schemes are applied, the results can of course be improved.

**Table 4.** Performance results of other publicly available finger vein datasets, ordered by publication year, "-" means that this information is not available. The "cla" column indicates contactless acquisition.


Regarding contactless hand vein recognition, Michael et al. [22] report an EER of 0.71% using palm veins only. Zhang et al. [23] only evaluate the KC value as measure of the registration between two feature vectors and thus, as an indicator for the recognition performance but they did not evaluate the actual recognition performance. The highest KC value they achieved was 1.1039. Fletcher et al. [24] reported an EER of 6.3% for their fully contactless hand vein based system for health patient identification. Table 5 summarises the achieved recognition performance for several publicly available hand vein recognition datasets, where the last row corresponds to the best results we achieved on our proposed contactless one so far. Note that, except for our proposed dataset and the PROTECT Mobile HandVein [25] dataset, all datasets have been acquired in a non contactless way. In the light of that and taking into account that we used only simple but well-established vein recognition schemes, the achieved recognition performance on our dataset is clearly competitive with other results reported in the literature. It is more than ten times better than the results published for the PROTECT Mobile HandVein [25] dataset and the results published by Fletcher et al. [24], and still twice as good as the results reported by Michael et al. [22], even though the only kind of hand normalisation we applied was the adaptive ROI extraction, correcting different scales, that is, different distances between the camera and the hand. No further tilt or non-planar rotation correction was applied. Again, note that our performance results should only serve as a baseline and can of course be improved.


**Table 5.** Related performance results of publicly available hand vein recognition datasets, ordered by publication year. The "cla" column indicates contactless acquisition.

While the achieved baseline results for the contactless hand vein dataset are quite competitive compared to the contactbased hand vein datasets, the results for the contactless finger vein datasets are clearly inferior to the ones that can be achieved for contactbased finger vein recognition. Contactless finger vein recognition is more challenging than contactless hand vein one for several reason. One reason is that the finger has a much smaller area than the palm of the hand. Thus, small misplacements can lead to a reduced visibility of the vein patterns and a reduced image quality in general, making the recognition more difficult. Moreover, the vascular pattern structure within the finger is more fine-grained than within the palm of the hand. Hence, tilts, rotation and bending of the finger have a more severe effect on the acquired images in terms of the resulting distortions present in the image, again leading to complications during the recognition process. These challenges have to be tackled by suitable normalisation and correction schemes in order to improve the recognition performance for the contactless finger vein data.

The sensor level fusion results clearly indicate that by combining different acquisition modes (finger vein, hand vein 850nm and hand vein 950 nm) the recognition performance can be considerably improved. By combining the hand vein images in the two different wavelengths, the average performance improvement over the best baseline one is 148%. By combining one finger and one hand sample, the best results (MC) are improved by 1183% and 171% over the baseline results for 850 nm and 950 nm hand veins, respectively. By combining all three modes the results were improved as well and are more than 8 times better than the best baseline one (MC). Hence, applying sensor level fusion is an easy way to further enhance the recognition performance of our capturing device.
