毕业生知识库

在这个信息爆炸的时代，人们对于保护个人信息及隐私的要求越来越高，更加安全且高效的身份认证技术自然成为了新的研究热点。同时在过去两年中，新冠疫情在全球范围内的席卷改变了人们的日常生活习惯。如今虽然疫情的高峰已经过去，但带来的影响深远，越来越多的人选择在生活中佩戴口罩，并且尽可能避免非必要的密切接触，给现今广泛使用的身份认证技术带来了诸多困难与挑战。面对这些困难和挑战，发展非接触、口罩遮挡场景下可用的身份认证技术是关键且急需解决的问题。眼部生物特征识别技术，特别是虹膜识别技术的非接触式采集和高精确性使其适用于应用场景与当下的时代背景。然而作为应用范围广泛、成熟完备的生物特征识别技术，应同时具备速度快、精度高、安全可靠、占用空间小的特点。眼部生物特征识别技术的进一步推广需要从高受控、中近距离场景拓展到低约束、远距离的更开放场景中进行研究。开放场景中往往包含有大量的不确定因素，包括但不限于：不同类型假体特征的恶意攻击、各种遮挡、复杂的环境光变化、纹理结构的形变、视线偏移等。这些因素往往会造成成像质量不稳定和整体下滑，引发图像预处理和特征识别环节的连锁反应，进而影响整个识别系统的性能。为了开发开放场景下鲁棒安全的眼部生物特征识别方法，本文针对上述挑战，从识别速度、识别精度、识别系统安全性、识别算法占用空间这四个角度出发展开研究。本文所取得的主要研究成果如下：

第一，提高虹膜识别系统的安全性和安全鲁棒性，解决复杂、混合假体虹膜攻击：本文提出了基于多层级分类的虹膜活体检测方法。对于常见的不同种类的假体虹膜，由于其假体部分特征分布不同，首次提出将不同种类的假体虹膜分为全局假体虹膜与局部假体虹膜，化繁为简。分别使用单独训练的两个深度学习网络从混合了不同种类的假体虹膜与真实虹膜的数据中分层级逐次分离出全局假体虹膜与局部假体虹膜，以达到活体检测的目标。该方法开创性地对假体虹膜定义了全新的分类方法，用分而治之的方法处理了虹膜活体检测中混合类型攻击的难题，提高了虹膜活体检测面对混合种类假体虹膜攻击时的安全性与鲁棒性。

第二，提高虹膜识别系统的识别精度与性能鲁棒性，增强对低质量虹膜图像的识别能力，从而提升虹膜识别系统的识别效率：本文提出了基于空域特征重建与多尺度局部特征的虹膜识别方法。相比于近距离、高配合度的虹膜识别系统，远距离、低配合度的虹膜识别系统所采集到的图像包含更多的低质量因子和非可控因素，针对低质量虹膜图像，提出了全新的对遮挡鲁棒、对形变鲁棒和无需对齐的端到端灵活识别框架，同时提取虹膜的全局特征与多尺度局部特征，设计了全新的虹膜特征表达与匹配方法，首次在虹膜识别领域引入了空域特征重建这一概念，以实现不同虹膜图像中局部区域之间的匹配能力。为了避免不同类图像中区域相似带来的误匹配情况，同时提取虹膜图像的全局特征作为相似度度量的一部分。该方法有效提升了中远距离、低配合度下虹膜识别的性能。

第三，提升虹膜识别系统的识别速度、节省虹膜识别算法占用的空间：本文提出了基于动量更新编码器和动态队列的虹膜识别方法。对于识别系统的现实应用来说，识别效率与计算速度是一个不可忽略的问题。虽然本文提出的方法有效提升了中远距离、低配合度下虹膜识别的性能，但其计算速度较慢是不容忽视的缺陷。在模型训练阶段，采用对编码器进行动量更新的策略，动态调整匹配字典队列，使队列大小不再受训练批量（batch）大小的影响，减少了参数量，降低虹膜识别系统的计算时间与存储成本。既保证了识别方法对于低质量图像的识别性能，又大大提升了识别效率，在识别精度与识别效率间达到了很好的平衡。

第四，进一步补足虹膜识别的劣势，利用眼部其他生物特征的互补性，扩大眼部生物特征识别系统的应用范围：本文提出了基于空域特征重建与无监督虹膜图像质量评价的眼部生物特征识别方法，选择采集虹膜时可以同时获得的眼周区域作为补充模态，并通过统一的识别框架对虹膜与眼周特征同时进行特征提取与匹配，在分数层对两种模态的匹配分数基于虹膜质量进行融合。相比于目前主流的基于人类主观视觉判断的虹膜质量评价标准，虹膜质量评价的目的应该是协助判断该虹膜图像是否适合用于识别算法，因此提出将虹膜图像质量重新定义为基于该图像的多尺度深度映射的相对鲁棒性。该虹膜质量评价方法旨在从识别任务的角度评估虹膜图像质量，进行更客观准确的质量度量。在这种虹膜质量评价方法的度量下，以虹膜质量评价分数作为分数层融合的权重参考。该方法集合了不同模态的优势、突破单一模态的限制，提高了整个识别系统的综合性能。

In this era of information explosion, people's demand for protecting personal information and privacy is increasing. Naturally, more secure and efficient identity authentication technology has become a new research hotspot. Meanwhile, the COVID-19 pandemic has brought profound changes in people's daily lives around the world in past two years. Although the peak of the pandemic has passed, its impact is far-reaching. More and more people decide to wear masks in their daily lives and avoid unnecessary close contact, which brings numbers of difficulties and challenges to the widely used identity authentication technology today. Facing these difficulties and challenges, it is is a key and urgent problem to develop non-contact and mask-robust identity authentication technology which can be utilized in various scenarios. Ocular biometric recognition technology, especially iris recognition technology, is suitable for application scenarios and the current background by its non-contact acquisition and high accuracy. However, as a widely used and mature biometric recognition technology, it should have the characteristics of fast speed, high accuracy, security, reliability, and small storage space. The further promotion of ocular biometric recognition technology needs to be extended from high-controlled, medium-to-short-distance scenarios to low-constrained, long-distance, and more open scenarios. Open scenarios often contain a large number of uncertain factors, including malicious attacks of different types of fake iris, various occlusions, complex environmental light changes, deformation of texture structure, and line-of-sight deviation. These factors will induce unstable image quality, image quality decline, triggering a chain reaction in the image preprocessing and feature recognition stages, which in turn affects the performance of the entire recognition system. In order to develop a robust and secure ocular biometric recognition method in open scenarios, the research is conducted from four perspectives: recognition speed, recognition accuracy, recognition system security, and recognition algorithm storage space. The main contributions of this paper are as follows:

Firstly, to enhance the security and robustness of iris recognition systems and to protect complex mixed-fake iris attacks, this paper proposes a hierarchical multi-class iris classification (HMC) method for liveness detection. Due to the different distribution of fake characteristics in various types of fake iris, the paper introduces a new classification method of fake iris into global fake iris and local fake iris, simplifying the problem. Distinct deep learning networks with independent training are used to gradually separate global and local fake iris from the mixed datasets that include different types of fake and real iris, achieving the goal of liveness detection. This method defines a novel classification approach for fake iris. It solves the challenge of mixed-type attacks in iris liveness detection by dividing and conquering. This method improves the security and robustness of iris liveness detection.

Secondly, to improve the recognition accuracy and robustness of iris recognition systems and enhance their ability to recognize low-quality iris images, this paper proposes an iris recognition method based on spatial feature reconstruction and multi-scale local features. Compared with iris recognition systems at close range with high cooperation, those at long distance and low cooperation capture images with more low-quality and uncontrollable factors. To recognize low-quality iris images, a new end-to-end flexible recognition framework which is robust to occlusion and deformation and does not require alignment is proposed. The global features and multi-scale local features of the iris are extracted, and a novel iris feature expression and matching method is designed. It is the first time that spatial feature reconstruction is introduced in iris recognition to achieve matching between local regions in different iris images. To avoid the misalignment caused by the similarity of regions in different eyes, global features are also extracted as part of the similarity measure. This method effectively improves the performance of iris recognition at long distances with low cooperation.

Thirdly, to improve the recognition speed of iris recognition systems and save the storage space, this paper proposes an iris recognition method training by momentum update encoder and dynamic queue. For practical applications of recognition systems, recognition efficiency and calculation speed are a major problem. Although the method proposed in this paper effectively improves the performance of iris recognition at long distances with low cooperation, its slow calculation speed is a significant drawback that cannot be ignored. In the training stage, a strategy of momentum updating the encoder is adopted, and the matching dictionary queue is dynamically adjusted, so that the queue size is no longer affected by the training batch size. This design reduces the number of parameters and lowers the calculation time and storage cost of the iris recognition system. The recognition method ensures recognition performance of low-quality images while greatly enhancing recognition efficiency, achieving a good balance between recognition accuracy and efficiency.

Fourthly, to further supply the shortcomings of iris recognition and expand the application scope of ocular biometric recognition systems by leveraging the complementarity of other biometric features in the eye, this paper proposes an ocular biometric recognition method based on spatial feature reconstruction and unsupervised iris image quality assessment. The periocular regions, which can be simultaneously acquired during iris acquisition, are selected as supplementary modalities. Iris and periocular features are contemporaneously extracted and matched by a unified recognition framework which fuses the matching scores of two modalities based on iris quality in a score-level fusion strategy. Compared with the current mainstream subjective human visual judgment-based iris quality assessment criteria, the purpose of iris quality assessment should be to assist in determining whether the iris image is suitable for use in recognition algorithms.Therefore, The image quality is defined as the magnitude of variations, which are calculated by the variations of embeddings generated from random subnetworks of cross-scale spatial feature extractor. This method aims to evaluate iris image quality from the perspective of recognition tasks and perform more objective and accurate quality measurements. By utilizing this iris quality assessment method, iris quality assessment scores are employed as weight references for score-level fusion. This method combines the advantages of different modalities and overcomes the limitations of single modality, improving the overall performance of the ocular recognition system.