Authors: Olotuah Roseline Temilade, Professor Oluwadare Samuel
Abstract: Accurate iris edge detection and boundary localization is seen as the prerequisite for high-performance biometric authentication systems. Traditional edge detection operators like, Canny, Sobel, Prewitt and feature-engineered shallow machine learning models are often beset with limitations such as severe sensitivity to noise, specular reflections, eyelid occlusions, and acute class imbalance when predicting thin boundary pixels. This paper presents a novel deep learning-driven ensemble framework that reformulates iris edge detection as a region-based binary semantic segmentation task using 2,639 high-resolution grayscale images (320 × 280 pixels, from the Standardized CASIA-Iris Interval benchmark database. The study adopts a Design Science Research (DSR) methodology, structured into three sequential phases: automated boundary localization for pseudo ground-truth generation using Circular Hough Transform, deep convolutional neural network segmentation (U-Net, ResNet-UNet, and Attention U-Net) and advanced ensemble mechanisms which includes Soft Voting, Stacking Meta-Learning, and Hybrid Fusion incorporating classical gradient maps. Raw grayscale iris images were made to undergo spatial standardization, Gaussian smoothing, and intensity normalization. The preprocessed tensors are passed concurrently through standalone CNN models and classical edge pipelines. The probability maps generated by CNNs are aggregated via soft voting, stacking, and hybrid fusion, and subsequently fused with classical gradient maps to derive precise binary segmentation masks and sharp boundary edges for improved robustness. The comparative evaluation, was done using four performance metrics; Intersection over Union (IoU), F1-Score, Precision and Recall. Experimental results demonstrate that deep learning architectures significantly outperform classical edge operators. Furthermore, ensemble models achieve superior segmentation dependability and boundary precision, attaining an Intersection over Union (IoU) of 0.9412, F1-Score of 0.9697, Precision of 0.9723, and Recall of 0.9671. The proposed region-based ensemble paradigm effectively allays pixel class imbalance, demonstrates superior robustness against intra-class variations, and establishes an accurate benchmark, especially, for next-generation biometric identification pipelines.
DOI: https://doi.org/10.5281/zenodo.24134661
