Abstract
Efficient deep palmprint recognition has become an urgent issue for the demand of personal identification on mobile/wearable devices. Compared to other biometrics, palmprint recognition has many unique advantages, e.g. richness of features, high user-friendliness, suitability for private security, etc. Existing deep learning based methods are computationally exhaustive in feature representation and learning, which are not suitable for large-scale deployment in portable authentication systems. In this paper, we combine hash coding and knowledge distillation to explore efficient deep palmprint recognition. Based on deep hashing network, palmprint images were converted to binary codes to save storage space and speed up matching. Combining hashing coding with knowledge distillation can further compress deep model to achieve an efficient recognition by light networks. Unlike previous palmprint recognition on datasets collected by dedicated devices in a controlled environment, we establish a novel database for unconstrained palmprint recognition, which consists of more than 30,000 images collected by 5 different mobile phones. Moreover, we manually labeled 14 key points on each image for region of interest (ROI) extraction. Comprehensive experiments were conducted on this palmprint database. The results indicate the feasibility of our database and the potential of palmprint recognition to be used as an efficient biometrics for deployment on consumer devices.
Introduction
Palmprint recognition offers a compelling biometric alternative due to its rich feature set and high privacy. However, deploying state-of-the-art deep learning models on mobile and wearable devices remains a challenge due to their high computational and memory requirements. Furthermore, most existing research utilizes datasets captured in controlled environments with specialized hardware, which does not reflect real-world mobile usage scenarios characterized by varying lighting and hand postures.
This paper addresses these issues by introducing an efficient recognition framework that combines Deep Hashing Networks (DHN) with Knowledge Distillation (KD). Additionally, it introduces the Xi’an Jiaotong University Unconstrained Palmprint (XJTU-UP) database to benchmark performance in realistic conditions.
Methodology
Deep Hashing Network (DHN)
To optimize matching speed and storage, the proposed system utilizes a Deep Hashing Network. Rather than generating high-dimensional floating-point feature vectors, the network learns to map palmprint images into compact 128-bit binary codes. This allows identity matching to be performed via simple XOR operations to calculate Hamming distances, a process that is natively efficient on mobile processors.
Knowledge Distillation (KD)
To achieve a lightweight model suitable for mobile deployment, the authors employ a Teacher-Student distillation framework. A pre-trained VGG-16 model serves as the "Teacher," possessing high representational power but a massive parameter count (156M). A "Student" network, significantly smaller with only 0.51M parameters (comprising 2 convolutional and 3 fully connected layers), is trained to mimic the Teacher's outputs.
ROI Extraction and Preprocessing
Handling unconstrained backgrounds requires robust preprocessing. The authors developed a pipeline for Region of Interest (ROI) extraction. By manually labeling 14 key points on palm images, they trained a regression tree algorithm to automatically locate these points and crop the central palm area, ensuring consistent input for the neural networks regardless of the original image's background or hand orientation.
The XJTU-UP Database
A significant contribution of this work is the XJTU-UP database, designed for unconstrained recognition. It contains over 30,000 images from 100 volunteers, captured using five different smartphone models (iPhone 6S, HUAWEI Mate8, LG G4, Galaxy Note5, and MI8). The dataset covers ten distinct domains, including variations in natural light and flash illumination, providing a rigorous test for algorithmic robustness.
Results and Discussion
The experimental results highlight the efficiency of the distilled student model. While a small network trained in isolation achieved only 45.2% accuracy, the distilled student reached 85.8% accuracy, demonstrating that distillation effectively transfers complex feature knowledge to simpler architectures.
In terms of efficiency, the student model achieved a 4x speedup in feature extraction, taking only 3.32 ms per image compared to the teacher's 12.30 ms. The use of flash was found to significantly improve performance by providing uniform illumination, which simplified the feature extraction task compared to natural light conditions.
Conclusion
This research demonstrates that combining deep hashing with knowledge distillation enables efficient palmprint recognition on mobile hardware. While an accuracy gap remains between the teacher and student models, the proposed framework significantly improves the performance of lightweight networks. Future work will focus on narrowing this gap and evaluating the system's real-world impact on battery life and latency on embedded devices.
Summary updated