Ocular biometrics usually includes iris recognition, sclera recognition, periocular recognition, retina recognition, eye movements recognition, etc., where iris recognition has been considered as one of the most stable, accurate and reliable biometric identification technologies and has been extensively studied. So far, iris recognition has been gradually developed from controlled scenarios to complex and less-constrained scenarios, e.g., at-a-distance, on-the-move, visible light illumination and using mobile devices. As a result, imaging devices often capture many low-quality iris images, which pose a serious challenge to iris recognition. In addition to the iris, a complete sclera region is usually contained in the captured iris image. Therefore, the accuracy and robustness of ocular recognition in complex scenarios are expected to be further improved by fusing
the iris and sclera traits.
Ocular image preprocessing is a key scientific problem in ocular biometrics, which not only defines the content of eye feature extraction and matching, but also performs the normalization and correction to images to reduce possible intra-class differences, hence it directly affects the overall recognition performance. In this thesis, we focus on the two main ocular biometrics, i.e., iris biometrics and sclera biometrics, with the aim of producing accurate and robust ocular preprocessing results for both unimodal iris/sclera recognition and multimodal iris-sclera fusion recognition. Specifically, we study three key preprocessing operations in complex scenarios, i.e., iris segmentation, iris normalization and sclera segmentation. Several robust, fast and accurate ocular image preprocessing methods are finally proposed for dealing well with adverse noise factors, e.g., occlusion, illumination variations, specular reflections, rotation, motion blur and off-angle. Our main contributions are summarized as follows:
1. In response to the lack of a large-scale iris segmentation benchmark dataset in complex scenarios, multiple iris datasets of various illumination (NIR, VIS), imaging sensors (professional iris camera, ordinary camera, mobile iris camera), imaging distances (close-range, long-range), races (white, Asian, black), user cooperation levels, and noise factors (out-of-focus, gaze deviation, specular reflections, motion blur, occlusions, dark iris, absence of iris, iris rotation, etc.) are firstly collected. Furthermore, a simple and efficient iris segmentation annotation method based on interactive ellipse and NURBS curve drawing operators is developed, which discards the previous cumbersome and coarse key points-based fitting method, and combines a newly proposed iris mask fine-tuning software IrisLabel V0.0 for the annotation of iris mask and iris inner/outer boundary in the benchmark dataset. Besides, a complete and comprehensive evaluation protocol is built to evaluate the accuracy, robustness, and generalization of iris segmentation, localization, and recognition as well as the model complexity. The proposed benchmark dataset has laid a solid data foundation for the following-up work and is beneficial for the development of more advanced iris segmentation methods in complex scenarios.
2. In response to the lack of a robust iris inner and outer boundaries localization technique for the existing deep learning based iris segmentation methods, an end-to-end multi-task learning and attention mechanism based iris segmentation model is proposed. It simultaneously predicts the iris mask, iris outer boundary and pupil mask for each image, and learns the iris texture category and boundary shape information from a large number of labeled datas. Furthermore, by exploiting the spatial priori constraints in the iris region, a simple yet effective post-processing method is proposed to achieve accurate and robust localization of parameterized iris inner and outer boundaries and refinement of predicted iris mask, which lays a good foundation for iris recognition in complex scenarios.
3. In response to the problems of heavy parameters, slow running speed and difficult prediction of iris outer boundary in the former model, an end-to-end lightweight iris segmentation model based on multi-label learning is proposed, and a knowledge distillation strategy is adopted to improve the segmentation and localization performance of the lightweight model. In this method, iris outer boundary is first filled up to alleviate of the positive/negative imbalance problem in the previous iris outer boundary prediction, and thus collectively forms a multi-label learning problem with iris mask and pupil mask. To achieve the multi-label learning, the DeepLabV3 model is adopted, and the cumbersome but superior teacher model and the lightweight yet sub-optimal student model are therefore derived. Finally, a knowledge distillation strategy is used to transfer the knowledge from the teacher network to the student network, eventually improving the performance of the student network. The proposed lightweight model is accurate and robust, and provides a feasible solution to iris segmentation in iris biometrics.
4. Based on Daugman's Rubber-Sheet model, a spatial transformer network based differentiable iris normalization model is proposed, and several normalization types are considered. The proposed model enables the normalized iris mask and normalized iris image to be generated within the network, which provides an accurate iris ROI region for subsequent iris feature analysis. Besides, it can also use the standard back propagation algorithm for end-to-end training, which lays a good foundation for the final end-to-end iris recognition system.
5. In response to the problems of the poor accuracy and low robustness in traditional model-driven methods and current deep learning-based methods when dealing with noisy sclera images, an attention assisted U-Net model is proposed to solve these challenges. Several different types of attention modules in channel-wise and spatialwise are incorporated into the central bottleneck part or skip connection part of the original U-Net, helping the new model implicitly learn to suppress irrelevant regions while highlighting salient features which are useful for sclera segmentation. Enjoying these benefits, the proposed model effectively improves the robustness and accuracy of
sclera segmentation, and wins the Sclera Segmentation Benchmarking Competition in Cross-resolution Environment (SSBC 2019).
In summary, the thesis systematically and deeply studies iris segmentation, iris normalization and sclera segmentation for ocular biometrics. A number of attempts are made and the accuracy and robustness of personal identification in complex scenarios
are significantly improved.