Introduction
Most bimodal biometric systems, such as those combining face and iris scans, often require multiple sensors or distinct capture steps. This complexity can be inconvenient for users and computationally taxing. Furthermore, unimodal systems like hand-shape recognition are susceptible to spoofing with artificial replicas. This paper proposes a bimodal system that extracts both geometry (shape) and texture (palmprint) from a single "peg-free" hand image captured by a digital camera. A key focus of the research is addressing feature redundancy to create models that are both accurate and efficient for real-time applications.
Methodology
The proposed system follows a multi-stage pipeline to process raw images into identifiable feature sets, as illustrated in the overall architecture below.
Image Preprocessing and ROI Extraction
The system first binarizes the hand image and employs a "best-fitting ellipse" method to align the hand vertically, ensuring rotation invariance. A distance transform is then used to automatically locate the palm center for region of interest (ROI) segmentation.
Feature Extraction
Two distinct types of features are extracted from the processed image:
- Palmprint Features (Texture): The Discrete Cosine Transform (DCT) is applied to overlapping blocks of the palm. This method is computationally efficient and potentially supported directly by camera hardware. Only the most significant 12.5% of coefficients are retained.
- Hand-Shape Features (Geometry): A set of 23 features is used, including 16 standard geometric measurements (finger lengths and widths) and 7 additional features such as perimeter, solidity, and "extent" to better characterize the hand's silhouette.
Fusion and Selection
Unlike systems that combine scores at the decision level, this approach uses feature-level fusion by concatenating raw feature vectors. To manage the resulting high-dimensional data, Correlation-based Feature Selection (CFS) is applied. CFS identifies features that are highly correlated with identity but uncorrelated with each other, effectively removing noise and redundant data.
Experimental Setup
The system was evaluated using an internal dataset of 1,000 images from 100 subjects (10 images per subject). To test temporal stability, images were collected in two sessions three months apart. The setup was "peg-free," allowing for a more natural user experience while increasing the complexity of image processing compared to traditional fixed-position systems.
Results and Discussion
The integration of multiple modalities significantly improved performance. While hand shape alone achieved ~89% accuracy and palmprint alone reached ~95%, their fusion pushed the recognition rate to nearly 98%.
The feature selection process proved vital for efficiency. For palmprints, over 52% of features (reducing from 144 to 69) could be discarded without loss of accuracy. In some instances, removing redundant features actually improved performance by mitigating overfitting. The best overall performance was achieved using a Support Vector Machine (SVM) classifier, reaching a recognition rate of 97.8% with only 75 selected features. Logistic Model Trees (LMT) also showed strong performance compared to simpler models like Naive Bayes.
Conclusion
This research demonstrates that combining hand geometry and palm texture from a single image provides a robust and efficient biometric solution. The use of DCT for texture and CFS for dimensionality reduction ensures the system is suitable for real-time hardware implementation. Future work will focus on scaling the dataset and improving robustness to varying environmental conditions and hand cleanliness.
Summary updated