Fast and Accurate ROI Extraction for Non-Contact Dorsal Hand Vein Detection in Complex Backgrounds Based on Improved U-Net

Rongwen Zhang, Xiangqun Zou, Xiaoling Deng, Ziyang Wang, Yifan Chen, Chengrui Lin, Hongxin Xing, Fen Dai2023 · DOI

TL;DR

This paper introduces an improved U-Net model for accurately and efficiently extracting the region of interest (ROI) from dorsal hand vein images, even in complex backgrounds. By incorporating residual modules, using bilinear interpolation for upsampling, and employing a Jensen-Shannon divergence loss, the model achieves higher accuracy with a significantly smaller footprint, making it suitable for resource-constrained embedded systems.

What Problem Does This Paper Address?

Accurate identification of individuals is crucial for security and various applications. While biometric systems like fingerprint and facial recognition have limitations (e.g., contact requirements, mask usage), dorsal hand vein (DHV) recognition offers a contactless, secure, and stable alternative. A critical step in DHV recognition is extracting the Region of Interest (ROI) – the specific area of the hand containing the vein patterns. Traditional image processing methods struggle with this task when faced with complex backgrounds, varying lighting conditions, and unconstrained hand poses. Deep learning approaches, particularly U-Net, show promise, but existing models can be computationally expensive and prone to degradation when deepened. This paper aims to bridge this gap by developing a lightweight yet highly accurate model for DHV ROI extraction that performs well even in challenging, real-world scenarios.

Methods

The core of this work is an improved U-Net architecture. The authors made several key modifications:

  • Residual Modules in Downsampling: To combat the "network degradation problem" that arises when deepening convolutional neural networks, residual modules (inspired by ResNet) were added to the downsampling path of the U-Net. This allows gradients to flow more easily, improving feature extraction.
  • Bilinear Interpolation for Upsampling: The original U-Net uses transpose convolution for upsampling, which can lead to "checkerboard artifacts." This paper replaces transpose convolution with bilinear interpolation, a more common and effective method for upscaling images, which also helps in fusing low-level and high-level features.
  • Jensen-Shannon (JS) Divergence Loss: Instead of solely relying on Mean Squared Error (MSE) for training, the authors introduced JS divergence loss. This loss function encourages the output feature map to approximate a Gaussian distribution, addressing the "multi-peak problem" where the model might predict multiple potential keypoint locations.
  • Soft-Argmax for Keypoint Decoding: To enable end-to-end training, Soft-Argmax is used to decode the keypoint coordinates from the final feature map. This is crucial because the standard argmax operation is non-differentiable, preventing gradient propagation during training.

These modifications collectively aim to improve accuracy, robustness, and efficiency.

Datasets

The study utilized two datasets:

  • Self-Built Dataset: Acquired using a custom infrared camera setup at South China Agricultural University. This dataset features unconstrained hand positions and backgrounds, including clean backgrounds, fluorescent lighting, clothing patterns, incomplete occlusion, and subjects wearing rings. It comprises 1890 images from 189 individuals.
  • Jilin University Public Dataset: This dataset consists of 2220 dorsal hand vein images from 222 people, with hands placed on a clean black platform. Images have a lower resolution (640x480 pixels) and typically feature the hand centered.

The self-built dataset is particularly valuable for its diversity and challenging conditions, which are more representative of real-world applications.

Key Insights and Ideas

  • Residual Connections are Key for Depth: The paper strongly suggests that for deeper networks, residual connections are not just an optimization but a necessity to prevent performance degradation and ensure effective feature learning.
  • Loss Function Matters for Distribution: The choice of loss function significantly impacts the model's output. JS divergence loss, by encouraging a Gaussian-like distribution in the heatmaps, helps in producing more precise and single-peaked predictions for keypoints.
  • Lightweight Models Can Be Powerful: The significant reduction in model size (from 35.0 M to 1.16 M) while increasing accuracy highlights the potential for deploying sophisticated deep learning models on embedded systems. This is a major step towards practical, on-device biometric solutions.
  • Dataset Diversity is Crucial: The performance drop when testing the self-built dataset-trained model on the Jilin University dataset (and vice-versa) underscores the importance of diverse training data. Combining datasets or using data augmentation strategies that mimic varied conditions is essential for generalization.

Key Results

  • Accuracy Improvement: The improved U-Net model achieved an accuracy of 98.6%, which is a 1% improvement over the original U-Net model.
  • Model Size Reduction: The improved model's file size is 1.16 M, a remarkable reduction to nearly 1/30th of the original U-Net model's size (35.0 M).
  • Efficiency: The improved model achieved a significantly lower Floating Point Operations per Second (FLOPs) of 939.0 M compared to the original U-Net's 7.44 G. Reasoning time also decreased from 146.12 ms to 59.57 ms.
  • Robustness: The model demonstrated high accuracy across various complex backgrounds and conditions, outperforming traditional image processing methods which failed completely in challenging scenarios.
  • Keypoint Detection: The model successfully located keypoints with high accuracy, enabling effective ROI extraction.

Limitations and Future Work

  • Uneven Illumination: The paper acknowledges that uneven illumination can still pose a challenge, leading to a higher Equal Error Rate (EER) of 5.81% in verification tasks.
  • Future Work: The authors plan to conduct in-depth research on the performance of feature vectors extracted from neural networks and traditional feature descriptors for dorsal hand vein ROI extraction. They also aim to further address the robustness issues caused by uneven illumination under unconstrained conditions.

Summary updated

BibTeX
@misc{zhang2023fast,
  author = {Rongwen Zhang and Xiangqun Zou and Xiaoling Deng and Ziyang Wang and Yifan Chen and Chengrui Lin and Hongxin Xing and Fen Dai},
  title = {{Fast and Accurate ROI Extraction for Non-Contact Dorsal Hand Vein Detection in Complex Backgrounds Based on Improved U-Net}},
  year = {2023},
  doi = {10.3390/s23104625},
}
Figure