Unconstrained Iris Recognition

Unconstrained Iris Recognition is a long-term research activity aimed at extending the accuracy and distinctiveness of iris biometrics beyond traditional, highly controlled acquisition systems. Conventional iris scanners generally require users to remain still, keep their eyes open, position themselves close to a dedicated sensor, and accept near-infrared illumination. The project instead studies ocular images captured under visible light, at variable distances, with ordinary cameras or smartphones, and from individuals who may be moving or unaware of the acquisition. It also considers facial portraits published on websites and social media, from which sufficiently detailed iris regions can sometimes be extracted automatically.

These less-constrained conditions introduce substantial image degradation. Typical problems include poor focus, motion blur, low spatial resolution, gaze deviation, partial occlusion by eyelids or eyelashes, heterogeneous reflections, shadows, weak iris–pupil contrast, compression artifacts, and large variations in illumination and sensor characteristics. The research therefore addresses the complete recognition pipeline: automatic ocular localisation, quality assessment, iris segmentation, removal of occlusions and reflections, texture normalisation, feature extraction, matching, and rejection of samples that do not contain sufficient biometric information.

Early contributions developed computational-intelligence and image-processing techniques for locating the iris and estimating its boundaries in non-ideal visible-light images. Subsequent work increasingly adopted deep neural networks to improve robustness. Dedicated encoder–decoder and U-Net-based architectures were designed to segment the iris and its reflection regions without assuming perfectly circular boundaries. Lightweight models were also considered so that processing could be performed on mobile or edge devices, limiting latency and reducing the need to transmit sensitive ocular images to remote servers.

A further research direction concerns automatic quality assessment. Deep models classify whether an ocular image is suitable for recognition by analysing the raw image together with information produced during iris localisation and segmentation. Removing unsuitable samples before matching can substantially improve recognition reliability: on I-SOCIAL-DB, the proposed quality-selection approach reduced the Equal Error Rate from 18.5% to 11.8%, while cross-dataset experiments demonstrated robustness to heterogeneous acquisition conditions.

The project also investigates super-resolution for web-sourced and social-media images, whose iris regions are often smaller than those normally required by biometric systems. Deep architectures such as RCAN, ESRGAN, and Real-ESRGAN are trained or adapted using high-resolution visible-light ocular datasets and then applied to challenging online images. Experiments showed that iris-specific training is more effective than directly using generic pretrained models and that improved visual quality does not always correspond to improved biometric accuracy. Nevertheless, dedicated super-resolution reduced recognition errors in controlled cross-resolution tests and produced smaller improvements on highly unconstrained web images, particularly for dark irises.

An important outcome of this research is I-SOCIAL-DB, a dataset containing 3,286 ocular regions extracted from 1,643 facial portraits of 400 subjects collected from websites and social media. The database includes manually annotated iris boundaries, occlusions, and reflections, enabling the evaluation of segmentation, quality assessment, recognition, and privacy-protection methods under genuinely uncontrolled conditions.

The ability to recognise individuals from publicly available portraits also creates significant privacy concerns. Iris patterns are highly stable over a person’s lifetime and may remain usable even when face recognition is hindered by age differences, masks, makeup, or other occlusions. The project therefore studies privacy-preserving countermeasures, including iris deidentification and synthetic texture replacement, to reduce recognisability while preserving the natural appearance and legitimate visual utility of facial images.

Overall, the research establishes a comprehensive framework for iris recognition in real-world conditions, combining robust artificial intelligence, image enhancement, quality control, lightweight edge processing, public datasets, and privacy-aware design. Its results are relevant to mobile authentication, border control, surveillance, forensic investigation, human–machine interaction, online-media analysis, and other applications in which accurate ocular recognition must operate beyond the constraints of conventional iris scanners.