TL;DR
This paper introduces a novel method for inner knuckle print (IKP) recognition that works reliably even with low-quality images from portable devices, varying lighting, and different hand poses. By developing new feature extraction and matching techniques, the system achieves state-of-the-art accuracy in challenging, real-world scenarios.
What Problem Does This Paper Address?
The increasing prevalence of portable devices with built-in cameras has fueled demand for convenient biometric authentication. However, using these devices for biometrics like fingerprints or palm prints is often hindered by uncontrolled lighting, low image quality, and variations in hand pose. This paper specifically targets Inner Knuckle Print (IKP) recognition, a biometric feature that is genetically determined and thus stable, but whose practical application has been limited by these environmental challenges. Existing methods often require controlled conditions or specialized hardware, making them unsuitable for everyday portable devices.
Methods
The authors propose a two-pronged approach to achieve illumination invariance and deformation tolerance:
- Illumination-Invariant Feature Extraction: They introduce a novel feature extraction scheme that aims to highlight the essential IKP structures while suppressing noise and artifacts introduced by varying lighting. This involves:
- Using the magnitude map of CompCode as a base feature map, which is known to tolerate monotonic illumination changes and small deformations.
- Applying a global feature extraction method based on spectral clustering to identify and preserve salient structures while eliminating unwanted ones.
- Refining the features using an empirical model that suppresses horizontal structures (often noise or less distinctive) and enhances vertical structures (more characteristic of IKP).
- Deformation-Tolerant Matching: To handle variations in hand orientation and perspective, they propose a Structure-Context Descriptor (SCD). This method:
- Divides the feature map into blocks and cells, summarizing pixel feature vectors within each cell to create a rough structure representation.
- Uses bilinear interpolation and a broadened cell region to mitigate aliasing issues from discrete block/cell structures.
- Employs the Earth Mover's Distance (EMD) for histogram matching, which is more robust than Euclidean distance in capturing true differences between feature histograms, especially under deformation.
Datasets
The authors built their own IKP database for benchmarking, as no public datasets were available.
- IKP Database: Collected from 100 volunteers (60 males, 40 females, aged 18-52).
- Each subject provided five images per hand.
- DB-1: Contains images captured under similar environments (2 images per hand), totaling 400 images and 1600 ROIs. This dataset is used to evaluate performance under controlled conditions.
- DB-2: Contains images captured under different environments with significant variations in posture and viewpoint (3 images per hand), totaling 800 images and 3200 ROIs. This dataset is used to evaluate robustness in uncontrolled environments.
Key Insights and Ideas
- IKP as a Promising Biometric: The paper reinforces the idea that IKP is a strong candidate for biometrics due to its genetic basis and clear structure, making it suitable even for low-quality images.
- The Challenge of Uncontrolled Environments: The core insight is that existing biometric methods often fail in real-world scenarios due to the combined effects of illumination variance and non-rigid deformations, which are inherent to portable device usage.
- Synergy of Local and Global Features: The method effectively combines local feature extraction (CompCode magnitude) with global structural analysis (spectral clustering) to achieve a more robust representation.
- EMD for Deformation Robustness: The adoption of Earth Mover's Distance for histogram comparison is a key innovation for handling the local deformations inherent in hand biometrics, outperforming traditional Euclidean distance.
- Empirical Refinement for Specific Structures: The use of an empirical model to specifically enhance vertical IKP structures and suppress horizontal ones demonstrates a nuanced understanding of the target biometric's characteristics.
Key Results
- Superior Accuracy in Uncontrolled Environments: On DB-2 (uncontrolled conditions), the proposed method significantly outperforms state-of-the-art methods.
- Integrated ROI Recognition: Achieved an Equal Error Rate (EER) of 0.74%, compared to the next best method (LGIC2) at 3.04%.
- Single ROI Recognition: Achieved an EER of 3.07%, significantly lower than existing methods (e.g., ImCompCode+MagCode at 11.23%).
- High Identification Rates: Achieved 100% Best Identification Rate (BIR) on DB-1 (controlled) and 93.5% on DB-2 (uncontrolled) for single ROI recognition.
- Deformation Tolerance: The system demonstrated robustness to significant changes in posture and viewpoint, as illustrated in Figure 14 and Figure 15.
- Illumination Tolerance: The method showed decent robustness to exposure variations (Figure 16) and different ambient light scenes (Figure 17).
- Efficiency: While slightly slower than some competitors on a PC, the method is efficient for matching due to offline feature extraction and parallel computation, with matching taking generally less than 50 ms for one thousand times.
Limitations and Future Work
- Computational Cost and Remote Server: The current algorithm still requires a remote server for execution due to its relatively high computational cost.
- Future Work: The authors believe that by redesigning and simplifying the algorithm, it can be made lightweight enough for direct execution on portable devices without a remote server.
- Threshold Determination: Determining a unified and reliable threshold for biometric systems remains a challenge. The paper uses a threshold based on minimum mismatching (FFRR) for stability, but this is an area for continued research.
- Database Limitations: While a new database was created, it's limited to five postures per hand. Future work could involve expanding the database to cover a wider range of postures for more comprehensive testing.
Summary updated