TL;DR
This paper investigates the effectiveness of three dimension reduction techniques (SVD, SDD, and FastMap) for iris recognition. The authors found that while SVD offers good quality, it's computationally expensive. FastMap provides a viable alternative for reducing dimensionality, and SDD shows potential despite some limitations in reconstruction quality.
What Problem Does This Paper Address?
Iris recognition is a promising biometric technology due to the unique and stable patterns of the iris. However, raw iris images, when represented as feature vectors, can be very high-dimensional. This "curse of dimensionality" poses challenges for efficient storage, processing, and retrieval. To combat this, dimension reduction techniques are employed to extract more compact and informative feature representations. This paper aims to compare the performance of several such techniques for iris recognition, specifically focusing on their impact on retrieval quality and the visual interpretability of the generated features.
Methods
The paper evaluates three dimension reduction methods:
- Singular Value Decomposition (SVD): A well-established technique that decomposes a matrix into three other matrices. In this context, it's used to find a lower-dimensional representation of the iris image collection. The authors discuss "rank-k SVD," which approximates the original matrix using only the top 'k' singular values and vectors.
- Semidiscrete Decomposition (SDD): An extension of SVD that aims to reduce storage requirements by constraining the entries of the decomposition matrices to {-1, 0, 1} and positive scalars for the diagonal matrix. This results in a more compressed representation but can lead to less accurate reconstructions.
- FastMap: A pivot-based dimension reduction technique designed for Euclidean spaces. It iteratively selects "pivots" (feature vectors) and projects other points onto the line defined by these pivots to reduce dimensionality. It's presented as a faster alternative to SVD for reducing dimensionality.
Datasets
The experiments were conducted on a real-life iris collection consisting of 384 irises. These were acquired using a TOPCON optical device and a CCD Sony camera, resulting in RGB images of size 576 × 768 pixels. For the experiments, only the red (R) component of the RGB image was used, as it was found to be more reliable. The dataset was preprocessed to exclude one iris per eye for querying, resulting in a collection of 256 irises from 64 people.
Key Insights and Ideas
- "Eigenirises" as Visual Features: The paper explores the concept of "eigenirises" (analogous to "eigenfaces" in face recognition), which are base vectors generated by dimension reduction methods. These eigenirises can offer insights into the common characteristics of the iris dataset.
- Trade-off between Compression and Reconstruction Quality: A key finding is the inherent trade-off between the degree of dimension reduction (and thus compression) and the quality of reconstructed images. Methods like SDD achieve high compression but at the cost of visual fidelity.
- SVD's Computational Cost: While SVD produces good results, its computational expense is highlighted as a significant drawback, especially for large collections or when frequent updates are needed.
- FastMap as a Practical Alternative: FastMap emerges as a promising method for dimensionality reduction when computational efficiency is a concern, offering a reasonable balance between performance and retrieval quality.
Key Results
- Retrieval Performance: Table 1 shows that for iris comparison using the Vector Space Model (VSM), SVD generally achieved the highest mean average precision (MAP) across different reduction dimensions (k). For k=128, SVD reached 49% MAP. FastMap also performed well, reaching 47% MAP at k=128, while SDD lagged significantly behind, achieving only 5% MAP.
- Identification Success: Table 2 indicates the number of queries where the correct person was identified. SVD consistently performed best, identifying the correct person in 84 out of 128 queries. FastMap was also competitive, with 96 successful identifications at k=128. SDD showed a much lower success rate, with only 5 identifications.
- Visual Quality: Reconstructed images from SVD were described as "somewhat blurred, but generally still recognizable." SDD reconstructions were noted as being "least exactly" generated, with a general loss of fine details, which is expected given its three-valued encoding.
Limitations and Future Work
- Iris Segmentation: The authors acknowledge that a critical limitation is the lack of robust iris segmentation, which involves removing the central piece (e.g., reflections), eyelids, and precisely identifying the iris position. They refer to existing work in this area.
- Computational Expense of SVD: The high computational cost of SVD is a major limitation, making it impractical for real-time applications or large-scale deployments without further optimization.
- Future Directions: The paper suggests exploring other dimension reduction methods like SparseMap, investigating faster pivot selection techniques for FastMap, and potentially leveraging Toeplitz matrices and their minimal eigenvalues for iris recognition.
Summary updated