Abstract: Biometrics based personal authentication is an effective way for automatically recognizing, with a high confidence, a person's identity. Recently, 3D ear shape has attracted tremendous interests in research field due to its richness of feature and ease of acquisition. However, the existing ICP (Iterative Closet Point)-based 3D ear matching methods prevalent in the literature are not quite efficient to cope with the one-to-many identification case. In this paper, we aim to fill this gap by proposing a novel effective fully automatic 3D ear identification system. We at first propose an accurate and efficient template-based ear detection method. By utilizing such a method, the extracted ear regions are represented in a common canonical coordinate system determined by the ear contour template, which facilitates much the following stages of feature extraction and classification. For each extracted 3D ear, a feature vector is generated as its representation by making use of a PCA-based local feature descriptor. At the stage of classification, we resort to the sparse representation based classification approach, which actually solves an $\ell_1$-minimization problem. To the best of our knowledge, this is the first work introducing the sparse representation framework into the field of 3D ear identification. Extensive experiments conducted on a benchmark dataset corroborate the effectiveness and efficiency of the proposed approach.
Introduction
Biometric identification—using unique physical traits like fingerprints or irises—is a cornerstone of modern security. The human ear is an excellent biometric candidate because its structure remains remarkably stable from age 8 to 70 and, unlike faces, it isn't affected by changing expressions.
While 3D ear scans provide rich geometric data that is immune to lighting changes, identifying a person from a large database (one-to-many identification) has historically been a computational nightmare. The industry standard, the Iterative Closest Point (ICP) algorithm, requires a probe scan to be mathematically "aligned" and compared to every single entry in a gallery one by one. As the database grows, the time required for identification becomes impractical for real-world use.
Proposed Methodology
The researchers propose a three-stage pipeline to move away from slow pairwise comparisons, transitioning from a geometric matching problem to a linear algebra problem.
1. Template-based Ear Detection
Instead of just finding the ear, the system locates the "ear pit" and aligns the entire scan to a canonical (standardized) coordinate system based on a pre-defined ear template. This is a critical step: because every ear is aligned to the same template at the start, they don't need to be aligned to each other later. The process involves first detecting the nose tip to define a search region, then identifying the ear pit within a specific angular sector.
2. PCA-based Feature Extraction
Once aligned, the 3D data is converted into a "feature map." The researchers used a local descriptor based on Principal Component Analysis (PCA) that captures the "unsymmetrical variation" (the unique bumps and ridges) of the ear's surface.
3. Sparse Representation-based Classification (SRC)
This is the core innovation. SRC treats the identification task as a massive puzzle. It tries to represent the "probe" ear as a sparse linear combination of all the ears in the gallery. By solving an $\ell_1$-minimization problem, the system can identify which person in the gallery most likely contributed to the probe's features in a single mathematical step.
Datasets and Experimental Setup
The study utilized the UND-J2 dataset from the University of Notre Dame. At the time of publication, this was the largest 3D ear dataset available, containing 2,436 side-profile 3D scans from 415 different individuals. To test the SRC framework effectively, the authors created three subsets where each person had at least 5, 7, or 10 samples in the gallery.
Key Results
The proposed system outperformed the traditional ICP method in both accuracy and, most dramatically, speed:
- Accuracy: On the largest test subset ($M=10$ samples per person), the SRC-based method achieved a 95.23% rank-1 recognition rate, compared to 94.09% for ICP.
- Speedup: For a gallery of 925 scans, the proposed algorithm took only 0.047 seconds per identification. In contrast, the ICP-based method took 144.31 seconds—a speedup of roughly 3,070 times.
- Detection: The automatic ear detection algorithm achieved a 90.87% success rate, significantly higher than previous automated methods on the same dataset.
Limitations and Future Work
Despite the performance gains, some limitations remain:
- Sample Requirements: SRC generally requires multiple samples per person in the gallery to build an effective "dictionary."
- Nose Dependency: The current ear detection pipeline relies on first locating the nose tip. If a 3D scan is cropped too tightly, detection might fail.
- Pose Sensitivity: Extreme head tilts or rotations beyond the template's range could still pose challenges.
Future work may involve exploring more robust local features or adapting the system for "in-the-wild" scans with non-ideal positioning.
Summary updated