Iris image segmentation for personal identification using unit gradient vectors and Iris-SAM

Iris recognition is one of a fundamental component of computer vision, extensively employed for authentication and identification within security and forensic contexts. Despite its broad utilization, the efficacy of current iris recognition systems is notably compromised under challenging conditions such as suboptimal lighting, partial iris visibility, and eye mis- alignment. These adverse circumstances frequently undermine the reliability of traditional iris recognition methodologies. Central to the effectiveness of iris recognition is the process of iris segmentation, which aims to precisely isolate the iris’s annular region from ocular images. Yet, the perfor- mance of conventional iris segmentation techniques typically deteriorates in diverse imaging environments. This thesis introduces an advanced approach to iris segmentation that inte- grates Unit Gradient Vectors (UGVs) with the Iris-Segment Anything Model (Iris-SAM) to enhance segmentation accuracy. This innovative methodology synergistically combines the Daugman algorithm with active contour methods, meticulously optimized using UGVs, to facilitate highly accurate initial segmentation. Further refinement is achieved through the implementation of Iris-SAM, which incorporates Focal Loss to address the significant class imbalance between iris and non-iris regions effectively. Comprehensive empirical evaluations conducted on the CASIA-Iris-Interval-v3 dataset affirm the superiority of this approach. The proposed model attains an average segmentation accuracy of 97.86%, surpassing the Iris-SAM benchmark of 96.94%. This improvement underscores the potential of combining traditional algorithms with modern machine learning techniques to overcome existing limitations and enhance the precision and robustness of iris recognition systems.

虹膜识别是计算机视觉的核心组成部分之一，被广泛应用于安全与法医场景中的身份验证与身份识别任务。尽管其应用范围广泛，但当前的虹膜识别系统在光照不佳、虹膜部分可见、眼球对齐偏差等挑战性场景下，其效能会显著受损。这些不利场景往往会削弱传统虹膜识别方法的可靠性。虹膜分割是影响虹膜识别效能的核心环节，其目标是从眼部图像中精准分割出虹膜的环状区域。然而，传统虹膜分割技术的性能在多样化的成像环境中通常会出现下降。本论文提出了一种先进的虹膜分割方法，该方法将单位梯度向量（Unit Gradient Vectors，UGVs）与虹膜分割一切模型（Iris-Segment Anything Model，Iris-SAM）相结合，以提升分割精度。该创新方法将道格拉斯曼算法（Daugman algorithm）与主动轮廓法（active contour methods）进行协同融合，并通过UGVs进行精细化优化，以实现高精度的初始分割。后续则通过引入Iris-SAM实现进一步的精细化处理，该模型结合了焦点损失（Focal Loss）以有效解决虹膜区域与非虹膜区域之间严重的类别不平衡问题。在CASIA-Iris-Interval-v3数据集上开展的全面实证评估证实了该方法的优越性。所提出的模型平均分割精度可达97.86%，优于Iris-SAM基准模型的96.94%。这一改进凸显了将传统算法与现代机器学习技术相结合的潜力，有助于克服现有局限，提升虹膜识别系统的精度与鲁棒性。