Multimodal biometric verification

This thesis presents a series of improvements to biometric verification systems focused on iris and periocular features, addressing key challenges like false positives and false negatives. First, we propose a segmentation framework that integrates the Segment Anything Model (SAM) with traditional Hough transform and active contour methods, improving segmentation accuracy. A novel initialization method for SAM and the use of full Hamming vectors in classification further enhance performance. Second, we extend traditional matching by combining Hamming distance, Jaccard distance, and Pearson correlation, and introduce a variance-based enrollment screening to filter out low-quality samples. Finally, we propose a score fusion approach that combines iris and periocular features, using an autoencoder and support vector classifier to boost verification accuracy. The periocular region, which includes the area around the eye such as eyelids, lashes, and skin texture, is less affected by occlusions and lighting variations than the iris. This makes it a reliable alternative, especially when the iris is partially visible or degraded. By integrating periocular cues with iris data, the system becomes more robust to challenging conditions.Evaluated on the CASIA-IrisV2 dataset, our combined methods significantly improve accuracy, F1-score, and recall, while reducing false negatives and enhancing robustness across varied input qualities.