Scalable iris image retrieval using attention-based CNNs and locality sensitive hashing (LSH)

Efficient and accurate retrieval of iris images from medium-scale databases plays an important role in national identification, access control, and identity verification. Traditional methods often struggle with scalability, memory demands, and latency as the dataset size grows. We present a scalable retrieval framework that employs a Convolutional Neural Network (CNN) enhanced with a Convolutional Block Attention Module (CBAM) to extract discriminative features. These features are organized via a clustering-based characteristic matrix and indexed using Min-Hash-based Locality-Sensitive Hashing (LSH) for efficient top-K retrieval. Evaluations on the CASIA-Iris-Interval. V3 and UBIRIS.V2 datasets show F1 scores of 93–97% and 95–96%, respectively, while maintaining sub-second query times and reducing index memory requirements by approximately 58% compared with dense-vector baselines. Although the LSH index requires more memory than highly compressed ITQ codes due to the additional Min-Hash signatures and hash-table structures, it provides a favorable memory–retrieval efficiency trade-off by substantially reducing search latency. Comparative evaluations show that LSH achieves its highest retrieval quality at an empirically optimized operating point (Top-K = 13), whereas ITQ and baseline approaches require substantially larger candidate sets (Top-K ≈ 400) to achieve comparable retrieval quality. These findings highlight the ability of LSH to maintain high retrieval accuracy while reducing candidate evaluation requirements and improving search efficiency. Extensive ablation studies and parameter analyses confirm the robustness and reproducibility of the proposed framework. These results suggest that the proposed LSH framework is a practical and effective solution for medium-scale iris retrieval under the evaluated experimental settings.