*1.1. Related Work*

It is important to note that, as far as it is known, there are no well-integrated and well-known commercial systems on the market based on the wrist vein modality. However, there are several studies in the research stage, as is exposed in the state-of-the-art of wrist Vascular Biometric Recognition (VBR) summarized in Table 1. It is divided into three units: dataset, capture device, and software algorithms. As it can be extracted from Table 1 and as far as it is known, there are only limited recent works, and there is only one public database for wrist vein modality: PUT [6] (50 subjects × 2 wrists × 4 samples × 3 sessions = 1200 images, 1100 genuine intraclass or mated comparisons and 108,900 impostor interclass or non-mated comparisons). This database is used in several works, e.g., [7], that also presents a complete and updated state-of-the-art of wrist VBR.

The rest of the works presented in Table 1 use two privately-distributed databases: UC3M [8] and Singapore [9]. Other less extensive works, e.g., [10], which are not presented in the table, use private databases collected with their own-designed system, as it is the case of the present study.

As far as is known, the cameras used mounted CCD sensors and LED type illumination with a wavelength of approximately 850 nm (considered the best near-infrared value for VBR).

It is essential to point out that all these databases require physical contact between the subject and the hardware part of the system, which reinforces the motivation discussed previously.

According to the recognition algorithms, all the studies follow the traditional recognition process against the trendy deep learning methods: preprocessing, feature extraction algorithm, and feature matching algorithm based on distances or machine learning techniques.

The process always begins with the preprocessing and enhancement of the near-infrared (NIR) images. The starting point is usually monochromatic images whose vein patterns are enhanced, for better definition and visualization, in the following order: contrast increase (e.g., histogram equalization), noise reduction (filters), binarization and skeletonization (e.g., Zhang and Suen [11]). Then, the task could continue (only in [9] in Table 1), with the extraction of the Region of Interest (ROI). For feature extraction, several techniques are applied: minutiae extraction, as the own algorithm discussed in [8]; feature extraction base on Local Binary Pattern (Dense Local Binary Pattern), [12]; Hessian matrix, [13]; and convolution approach, [14]. The matching algorithms are based on distance (Hausdorff distance [9,14] and own minutiae algorithm [8]) and cross-correlation comparison [13].

Traditional Machine Learning methods for matching are only employed in [12] (Support Vector Machines).

Computing time for the entire software algorithm is given in the latest works, [12,13], revealing the evolution of biometric systems nowadays. As the results of the proposed system indicate (Section 3, Section 3.2.2), computing time is a critical variable in the integration of real-time biometric systems.

The performance for all works indicated in Table 1, based on the Equal Error Rate (EER), varies between 0.14% and 2.27%. These values should be considered reduced enough, but it is important to remark, as it has been mentioned, all the devices required physical contact with the users, fixing the wrist position and easing the recognition task. The images obtained are extremely invariant in scale and orientation, which translate into really high biometric performance, as it is demonstrated in the current work, with entirely similar features extracted. In this sense, a new scale-orientation-invariant algorithm is presented in the current study.

Another important factor noticed in the capture devices, due to the contact feature, is the immunity to the environmental or external light. This light does not reach the capture device due to the closed space between the camera and the wrist. Again, as a result, the similarity between the images is improved and, of course, the recognition performance. A non-contact system, as it is demonstrated in this paper, is affected by the external light conditions despite the extra capture light illumination. These two factors are the goals to overcome in order to improve wrist VBR recognition and obtain contactless devices.


**Table 1.**Summary of the state-of-the-art for wrist Vascular Biometric Recognition (VBR).
