Performance Analysis of Handcrafted Features for Iris Recognition Using the IITD Dataset

Abstract

Biometric authentication has become an important part of modern cybersecurity systems because it provides a convenient and reliable means of verifying user identity. Among different biometric modalities, iris recognition is particularly attractive due to the distinct and stable characteristics of the iris. This study evaluates six widely used feature extraction techniques for iris recognition: Local Binary Patterns (LBP), Gabor Filters, Histogram of Oriented Gradients (HOG), Discrete Wavelet Transform (DWT), Scale-Invariant Feature Transform (SIFT), and a hybrid LBP–SIFT method. Experiments were conducted using the IITD Iris Image Database to assess the suitability of these techniques for secure biometric authentication. After normalization, iris images were processed using each feature descriptor, and the resulting features were compared using cosine distance. Performance was assessed through Receiver Operating Characteristic (ROC) curves, Area Under the Curve (AUC), and Equal Error Rate (EER). The results showed that LBP achieved the highest AUC, whereas SIFT produced the lowest EER, indicating strong recognition performance. In contrast, Gabor Filters showed comparatively weaker results, while HOG and DWT provided a reasonable balance between computational complexity and recognition accuracy. The hybrid LBP–SIFT approach produced only a marginal improvement over the individual descriptors. Overall, the findings suggest that handcrafted features can still provide practical solutions for biometric security, particularly in lightweight access-control applications where computational resources are limited and the use of deep learning models may not be practical.