Sanil Gangothri, Prakasha K Krishna, Prabhu Srikanth, Nayak Vinod, Jayakala Aparna
Information Communication Technology, Manipal Institute of Technology (MIT), Manipal Academy of Higher Education, Manipal, 576104, India.
Information Communication Technology, Manipal Academy of Higher Education (MAHE),, Manipal Academy of Higher Education, Manipal Institute of Technology (MIT), Manipal, 576104, India.
F1000Res. 2025 Jun 20;14:444. doi: 10.12688/f1000research.162911.2. eCollection 2025.
In computer vision and image processing, face recognition is increasingly popular field of research that identifies similar faces in a picture and assigns a suitable label. It is one of the desired detection techniques employed in forensics for criminal identification.
This study explores face recognition system for monozygotic twins utilizing three widely recognized feature descriptor algorithms: Scale-Invariant Feature Transform (SIFT), Speeded-Up Robust Features (SURF), and Oriented Fast and Rotated BRIEF (ORB)-with region-specific facial landmarks. These landmarks were extracted from 468 points detected through the MediaPipe framework, which enables simultaneous recognition of multiple faces. Quantitative similarity metrics t served as inputs for four classification methods: Support Vector Machine (SVM), eXtreme Gradient Boost (XGBoost), Light Gradient Boost Machine (LGBM), and Nearest Centroid (NC). The effectiveness of these algorithms was tested and validated using challenging ND Twins and 3D TEC datasets, the most difficult data sets for 2D and 3D face recognition research at Notre Dame University.
Testing with Notre Dame University's challenging ND Twins and 3D TEC datasets revealed significant performance differences. Results demonstrated that 2D facial images achieved notably higher recognition accuracy than 3D images. The 2D images produced accuracy of 88% (SVM), 83% (LGBM), 83% (XGBoost), and 79% (NC). In contrast, the 3D TEC dataset yielded a lower accuracy r of 74%, 72%, 72%, and 70%, with the same classifiers.
The hybrid feature extraction approach proved most effective, with maximum accuracy rates reaching 88% for 2D facial images and 74% for 3D facial images. This work contributes significantly to forensic science by enhancing the reliability of facial recognition systems when confronted with indistinguishable facial characteristics of monozygotic twins.
在计算机视觉和图像处理领域，人脸识别是一个越来越受欢迎的研究领域，它能识别图片中的相似面孔并赋予合适的标签。它是法医用于刑事鉴定的理想检测技术之一。
本研究利用三种广泛认可的特征描述符算法探索同卵双胞胎的人脸识别系统：尺度不变特征变换（SIFT）、加速稳健特征（SURF）和具有区域特定面部标志点的定向快速旋转BRIEF（ORB）。这些标志点是从通过MediaPipe框架检测到的468个点中提取的，该框架能够同时识别多张面孔。定量相似性指标作为四种分类方法的输入：支持向量机（SVM）、极端梯度提升（XGBoost）、轻量级梯度提升机（LGBM）和最近质心（NC）。使用具有挑战性的圣母大学双胞胎（ND Twins）和3D TEC数据集对这些算法的有效性进行了测试和验证，这是圣母大学二维和三维人脸识别研究中最具难度的数据集。
使用圣母大学具有挑战性的ND Twins和3D TEC数据集进行测试，结果显示出显著的性能差异。结果表明，二维面部图像的识别准确率明显高于三维图像。二维图像使用SVM的准确率为88%，LGBM为83%，XGBoost为83%，NC为79%。相比之下，3D TEC数据集使用相同分类器的准确率较低，分别为74%、72%、72%和70%。
混合特征提取方法被证明是最有效的，二维面部图像的最高准确率达到88%，三维面部图像为74%。这项工作通过提高人脸识别系统在面对同卵双胞胎难以区分的面部特征时的可靠性，为法医学做出了重大贡献。
