Region of interest (ROI) localization is one of the key preprocessing technologies for a finger-vein identification system, so an effective ROI definition can improve the matching accuracy. However, due to the impact of uneven illumination, equipment noise, as well as the distortion of finger position, etc., these make accurate ROI localization a very difficult task. To address these issues, in this paper, we propose a robust finger-vein ROI localization method, which is based on the 3 $\sigma$ criterion dynamic threshold strategy. The proposed method includes three main steps: First, the Kirsch edge detector is introduced to detect the horizontal-like edges in the acquired finger-vein image. Then, the obtained edge gradient image is divided into four parts: upper-left, upper-right, lower-left, and lower-right. For each part of the image, the three-level dynamic threshold, which is based on the 3 $\sigma$ criterion of the normal distribution, is imposed to obtain more distinct and complete edge information. Finally, through labeling the longest connected component at the same horizontal line, two reliable finger boundaries, which represent the upper and lower boundaries, respectively, are defined, and the ROI is localized in the region between these two boundaries. Extensive experiments are carried out on four different finger-vein image datasets, including three publicly available datasets and one of our newly developed finger-vein datasets with 37,080 finger-vein samples and 1030 individuals. The experimental results indicate that our proposed method has very competitive ROI localization performance, as well as satisfactory matching results on different datasets.

Finger vein (FV) biometrics is one of the most promising individual recognition traits, which has the capabilities of uniqueness, anti-forgery, and bio-assay, etc. However, due to the restricts of imaging environments, the acquired FV images are easily degraded to low-contrast, blur, as well as serious noise disturbance. Therefore, how to extract more efficient and robust features from these low-quality FV images, remains to be addressed. In this paper, a novel feature extraction method of FV images is presented, which combines curvature and radon-like features (RLF). First, an enhanced vein pattern image is obtained by calculating the mean curvature of each pixel in the original FV image. Then, a specific implementation of RLF is developed and performed on the previously obtained vein pattern image, which can effectively aggregate the dispersed spatial information around the vein structures, thus highlight vein patterns and suppress spurious non-boundary responses and noises. Finally, a smoother vein structure image is obtained for subsequent matching and verification. Compared with the existing curvature-based recognition methods, the proposed method can not only preserve the inherent vein patterns, but also eliminate most of the pseudo vein information, so as to restore more smoothing and genuine vein structure information. In order to assess the performance of our proposed RLF-based method, we conducted comprehensive experiments on three public FV databases and a self-built FV database (which contains 37,080 samples that derived from 1030 individuals). The experimental results denoted that RLF-based feature extraction method can obtain more complete and continuous vein patterns, as well as better recognition accuracy.

Finger vein (FV) biometrics have garnered considerable attention due to their inherent non-contact nature and high security, exhibiting tremendous potential in identity authentication and beyond. Nevertheless, challenges pertaining to the scarcity of training data and inconsistent image quality continue to impede the effectiveness of finger vein recognition (FVR) systems. To tackle these challenges, we introduce the visual feature-guided diamond convolutional network (dubbed ‘VF-DCN’), a uniquely configured multi-scale and multi-orientation convolutional neural network. The VF-DCN showcases three pivotal innovations: Firstly, it meticulously tunes the convolutional kernels through multi-scale Log-Gabor filters. Secondly, it implements a distinctive diamond-shaped convolutional kernel architecture inspired by human visual perception. This design intelligently allocates more orientational filters to medium scales, which inherently carry richer information. In contrast, at extreme scales, the use of orientational filters is minimized to simulate the natural blurring of objects at extreme focal lengths. Thirdly, the network boasts a deliberate three-layer configuration and fully unsupervised training process, prioritizing simplicity and optimal performance. Extensive experiments are conducted on four FV databases, including MMCBNU_6000, FV_USM, HKPU, and ZSC_FV. The experimental results reveal that VF-DCN achieves remarkable improvement with equal error rates (EERs) of $0.17\%$, $0.19\%$, $2.11\%$, and $0.65\%$, respectively, and Accuracy Rates (ACC) of $100\%$, $99.97\%$, $98.92\%$, and $99.36\%$, respectively. These results indicate that, compared with some existing FVR approaches, the proposed VF-DCN not only achieves notable recognition accuracy but also shows fewer number of parameters and lower model complexity. Moreover, VF-DCN exhibits superior robustness across diverse FV databases.

The evolution of deep learning has promoted the performance of finger vein verification systems, but also brings some new issues to be resolved, including high computational burden, massive training sample demand, as well as the adaptability and generalization to various image acquisition equipment, etc. In this paper, we propose a novel and lightweight network architecture for finger vein verification, which was constructed based on a Siamese framework and embedded with a pair of eight-layer tiny ResNets as the backbone branch network. Therefore, it can maintain good verification accuracy under the circumstance of a small-scale training set. Moreover, to further reduce the number of parameters, Gabor orientation filters (GoFs ) were introduced to modulate the conventional convolutional kernels, so that fewer convolutional kernels were required in the subsequent Gabor modulation, and multi-scale and orientation-insensitive kernels can be obtained simultaneously. The proposed Siamese network framework (Siamese Gabor residual network (SGRN)) embeds two parameter-sharing Gabor residual subnetworks (GRNs) for contrastive learning; the inputs are paired image samples (a reference image with a positive/negative image), and the outputs are the probabilities for accepting or rejecting. The subject-independent experiments were performed on two benchmark finger vein datasets, and the experimental results revealed that the proposed SGRN model can enhance inter-class discrepancy and intra-class similarity. Compared with some existing deep network models that have been applied to finger vein verification, our proposed SGRN achieved an ACC of $99.74\%$ and an EER of $0.50\%$ on the FV-USM dataset and an ACC of $99.55\%$ and an EER of $0.52\%$ on the MMCBNU_6000 dataset. In addition, the SGRN has smaller model parameters with only 0.21 $\times {10}^6$ Params and 1.92 $\times {10}^6$ FLOPs, outperforming some state-of-the-art FV verification models; therefore, it better facilitates the application of real-time finger vein verification.

At present, ResNet and DenseNet have achieved significant performance gains in the field of finger-vein biometric recognition, which is partially attributed to the dominant design of cross-layer skip connection. In this manner, features from multiple layers can be effectively aggregated to provide sufficient discriminant representation. Nevertheless, an over-dense connection pattern may induce channel expansion of feature maps and excessive memory consumption. To address these issues, we proposed a low memory overhead and fairly lightweight network architecture for finger-vein recognition. The core components of the proposed network are a sequence of sparsified densely connected blocks with symmetric structure. In each block, a novel connection cropping strategy is adopted to balance the channel ratio of input/output feature maps. Beyond this, to facilitate smaller model volume and faster convergence, we substitute the standard convolutional kernels with separable convolutional kernels and introduce a robust loss metric that is defined on the geodesic distance of angular space. Our proposed sparsified densely connected network with separable convolution (hereinafter dubbed ‘SC-SDCN’) has been tested on two benchmark finger-vein datasets, including the Multimedia Lab of Chonbuk National University (MMCBNU)and Finger Vein of Universiti Sains Malaysia (FV-USM), and the advantages of our SC-SDCN can be evident from the experimental results. Specifically, an equal error rate (EER) of 0.01% and an accuracy of 99.98% are obtained on the MMCBNU dataset, and an EER of 0.45% and an accuracy of 99.74% are obtained on the FV-USM dataset.

Currently, finger vein recognition (FVR) stands as a pioneering biometric technology, with convolutional neural networks (CNNs) and Transformers, among other advanced deep neural networks (DNNs), consistently pushing the boundaries of recognition accuracy. Nevertheless, these DNNs are inherently characterized by static, continuous-valued neuron activations, necessitating intricate network architectures and extensive parameter training to enhance performance. To address these challenges, we introduce an adaptive firing threshold-based spiking neural network (ATSNN) for FVR. ATSNN leverages discrete spike encodings to transforms static finger vein images into spike trains with spatio-temporal dynamic features. Initially, Gabor and difference of Gaussian (DoG) filters are employed to convert image pixel intensities into spike latency encodings. Subsequently, these spike encodings are fed into the ATSNN, where spiking features are extracted using biologically plausible local learning rules. Our proposed ATSNN dynamically adjusts the firing thresholds of neurons based on average potential tensors, thereby enabling adaptive modulation of the neuronal input-output response and enhancing network robustness. Ultimately, the spiking features with the earliest emission times are retained and utilized for classifier training via a support vector machine (SVM). Extensive experiments conducted across three benchmark finger vein datasets reveal that our ATSNN model not only achieves remarkable recognition accuracy but also excels in terms of reduced parameter count and model complexity, surpassing several existing FVR methods. Furthermore, the sparse and event-driven nature of our ATSNN renders it more biologically plausible compared to traditional DNNs.