Abstract
The ignored side effect reflecting in the introduction of mismatching brought by contrast enhancement in representative SIFT based vein recognition model is investigated. To take advantage of contrast enhancement in increasing keypoints generation, hierarchical keypoints selection and mismatching removal strategy is designed to obtain state-of-the-art recognition result.
Introduction
Hand-dorsa (back of the hand) vein recognition is a promising biometric because it is difficult to forge and works with non-contact sensors. However, near-infrared (NIR) vein images typically suffer from low contrast and "dark" regions, making it hard for local invariant descriptors like SIFT or SURF to find enough keypoints for reliable matching.
To address this, researchers commonly apply Contrast Enhancement (CE). While CE successfully generates more keypoints, it also amplifies noise and creates "phantom" features in non-vein areas. This paper introduces a Scale and Contrast Invariant Feature Matching (SCIFM) strategy that captures the "quantity" benefits of CE while eliminating the "quality" penalties.
Proposed Framework
The SCIFM framework follows a structured pipeline from image acquisition to final decision, integrating contrast enhancement and specialized feature extraction.
Methods
- Contrast Enhancement (CE): The authors evaluated 18 different CE methods (including Histogram Equalization and Laplacian sharpening) to balance feature generation.
- Feature Extraction: SIFT keypoints are extracted from the enhanced images.
- Hierarchical Mismatching Removal: This is the core innovation, consisting of two stages:
- Stage 1: Non-Vein Keypoint Removal: A "vein skeleton" is generated. Keypoints detected in non-vein background areas are discarded as noise.
- Stage 2: Mirror Matching Strategy: A "sanity check" where features must be more similar to the target image than to other points within the same source image.
Datasets
The researchers utilized the CUMT Hand-Dorsa Vein Image Database, containing 2,000 images from 100 individuals. Both left and right hands were captured under varying conditions (time of day, illumination) to ensure robustness.
Key Results and Insights
- The CE Paradox: Contrast enhancement helps SIFT identify more local extrema but increases sensitivity to noise. Using DHE (DoG filter and Histogram Equalization), the average number of keypoints jumped from 12 to 479.
- Performance: Without the proposed removal strategy, standard SIFT had an EER of 18.4%. With SCIFM and DHE enhancement, the EER dropped significantly to 0.061%.
- Benchmarking: This approach outperformed RootSIFT (0.105%) and LBP-based models (0.113%) on the same dataset.
Conclusion
The proposed SCIFM strategy effectively mitigates the noise introduced by contrast enhancement in vein recognition. By using spatial filtering (vein skeletons) and self-similarity checks (mirror matching), the system achieves state-of-the-art accuracy. Future work will explore integrating these hierarchical filters into deep learning architectures.
Summary updated