TL;DR
This paper introduces PalmRCNN, a novel Deep Convolutional Neural Network (DCNN) approach for contactless palmprint and palmvein recognition. To address the lack of large-scale, high-quality contactless palmvein datasets, the authors developed a new acquisition device and created such a dataset, which is now publicly available.
What Problem Does This Paper Address?
Biometric authentication is crucial for security and access control. While palmprint and palmvein recognition offer unique and stable identification features, existing methods often struggle with accuracy and robustness, particularly for contactless scenarios. A significant bottleneck has been the scarcity of large-scale, publicly available datasets of high-quality, contactless palmvein images, hindering the development and evaluation of advanced recognition algorithms. This paper aims to bridge this gap by proposing a powerful DCNN-based recognition system and a comprehensive new dataset.
Methods
The core of the proposed solution is PalmRCNN, a DCNN architecture adapted from Google's Inception-ResNet-v1. The authors modified this architecture to effectively extract deep and representative features from palmprint and palmvein images.
The PalmRCNN pipeline involves:
- Region of Interest (ROI) Extraction: A robust algorithm is used to align palmprint images by extracting a standardized 160x160 ROI, accounting for translation, rotation, and scaling variations.
- Modified Inception-ResNet-v1: This network, comprising a stem and several Inception-ResNet and Reduction modules, is trained to learn rich features. The architecture combines Inception's multi-scale processing with residual connections to enable deeper network training and improved performance. The final feature vector is 128-dimensional.
- Identification: For identification (one-to-many matching), a Support Vector Machine (SVM) classifier is trained using the extracted feature vectors.
- Verification: For verification (one-to-one matching), the Euclidean distance between feature vectors of two images is used to determine if they belong to the same individual.
Datasets
- Tongji University Contactless Palmprint Dataset: Used for training and evaluating the palmprint recognition component. It contains 12,000 images from 600 palms.
- New Large-Scale Contactless Palmvein Dataset: This is a key contribution of the paper. The authors developed a novel acquisition device and collected a dataset comprising 12,000 high-quality contactless palmvein images from 600 different palms. This dataset was collected over two sessions with a two-month interval and is made publicly available.
The paper also references and compares results against existing palmvein datasets like Kabacinski et al. [33], Hao et al. [34], and Ferrer et al. [35].
Key Insights and Ideas
- DCNNs for Palm Biometrics: The paper demonstrates that DCNNs, specifically adapted architectures like Inception-ResNet-v1, can effectively learn robust and discriminative features for palmprint and palmvein recognition, outperforming traditional low-level feature extraction methods.
- The Importance of Large-Scale Datasets: The authors highlight that the lack of a comprehensive, high-quality contactless palmvein dataset has been a major impediment. Their contribution of a new, large-scale dataset is crucial for advancing research in this area.
- User-Friendly Acquisition Device: The design of a user-friendly, ergonomic acquisition device that produces images with simple backgrounds is a practical innovation that simplifies preprocessing and improves overall system robustness.
- Combining Palmprint and Palmvein: While this paper focuses on each modality separately, the authors suggest that future work could explore fusing palmprint and palmvein information for even higher accuracy.
Key Results
- Palmprint Identification: The proposed PalmRCNN achieved 100.00% accuracy, precision, and recall on the Tongji University Contactless Palmprint Dataset, outperforming several state-of-the-art methods.
- Palmprint Verification: The method achieved an Equal Error Rate (EER) of 2.74%, which is lower than competing methods (e.g., 6.46%, 3.37%, 4.50%).
- Palmvein Identification: The system achieved 100.00% accuracy, precision, and recall on the newly collected palmvein dataset.
- Palmvein Verification: The method achieved an EER of 2.30%, outperforming existing methods (e.g., 4.00%, 2.36%).
Limitations and Future Work
- Limitations: The paper does not explicitly detail limitations of the PalmRCNN model itself, but the performance of DCNNs is inherently dependent on the quality and quantity of training data. The authors acknowledge that traditional methods have room for improvement.
- Future Work: The authors explicitly state their intention to explore combining palmprint and palmvein recognition in future work to further enhance recognition accuracy.
Summary updated