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

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Traditional image processing struggles with accurate region extraction from complex dorsal hand vein images.
• Non-contact vein imaging presents challenges due to background complexity and feature variability.

Purpose of the Study:

• To develop an improved U-Net model for precise dorsal hand keypoint detection.
• To enhance feature extraction and address model degradation in dorsal hand vein analysis.
• To create a computationally efficient model suitable for low-resource platforms.

Main Methods:

• An improved U-Net architecture incorporating a residual module in the downsampling path.
• Utilized Jensen-Shannon (JS) divergence loss function for improved feature map distribution.
• Employed Soft-argmax for end-to-end keypoint coordinate calculation.

Main Results:

• The improved U-Net model achieved an accuracy of 98.6%, surpassing the original U-Net by 1%.
• The model size was reduced to 1.16 M, indicating significantly fewer parameters.
• Demonstrated superior feature extraction and robustness against feature map multi-peak issues.

Conclusions:

• The proposed improved U-Net model effectively performs dorsal hand keypoint detection for non-contact vein images.
• The model's high accuracy and reduced size make it suitable for deployment on edge-embedded systems.
• This approach offers a practical solution for region of interest extraction in resource-constrained environments.