【作者】 刘婷婷；

【导师】 张忠波；

【作者基本信息】 吉林大学 ， 计算数学， 2022， 硕士

【摘要】 虹膜识别技术是基于眼睛中的虹膜进行身份识别的生物特征识别技术,相较指纹、人脸识别等其他生物特征识别方式,具有识别准确率更高、误识率更低、无需重复注册、非接触和极难伪造等优势,被认为是除DNA以外“最可靠的生物识别技术”。与其它生物识别技术相比,虹膜识别技术应用的最大的难点是在无需用户高度配合的情况下,实现远距离采集用户的清晰双目虹膜图像。与近距离虹膜采集相比,远距离虹膜采集无论从硬件到软件都迫切需要解决多项关键技术难题。人眼的虹膜直径为11毫米,普通的拍摄设备很难实现在2米左右的距离采集得到清晰虹膜纹理图像,特殊的成像系统、光源系统才能够解决这一问题。同时,要解决远距离虹膜采集带来的虹膜图像问题,例如,拍摄时聚焦不准导致图像离焦模糊、低分辨率虹膜图像,被拍摄者身体不稳定引起的运动模糊,眼睑、睫毛遮挡虹膜等问题,这些问题都是实现远距离虹膜识别的难点。采集到的虹膜图像中,低质量虹膜图像占比较大,从采集到的大量低质量虹膜图像(低分辩、离焦模糊、运动模糊)中获得高分辨清晰的虹膜图像是准确识别虹膜的前提条件,这就需要设计精准、高效的虹膜图像超分重建及模糊虹膜图像复原算法。图像超分辨(Image Super-Resolution)是指由低分辨率图像重建高分辨率图像。其中,单帧图像超分辨(Single Image Super-Resolution,SISR)的研究较为广泛,也出了很多成果,多帧超分辨(Multi-Frame Super-Resolution,MFSR)由于问题本身固有的难度,研究相对较少。深度学习在图像处理领域大规模应用以来,图像超分辨的质量有了质的提升。在远距离虹膜识别领域,由于光学原理及图像采集设备的限制,很多时候只能采集到低分率图像,而且图像质量对识别模型的性能影响非常大。这些促使我们使用深度学习的方法研究虹膜图像的超分辨重建问题。我们通过特定远距离虹膜图像采集设备,采集了涵盖不同模糊程度、不同分辨率、不同遮挡程度的4300张虹膜图像作为训练和测试的数据集。在前人的研究基础之上,结合远距离虹膜采集的条件和应用领域,本文构建了对齐(Alignment)—融合(Fusion)—重建(Restruction)的总体网络框架。多帧超分辨的输入是多个含有噪声、具有相对位移且无序的低分辨图像,图像具有未知的位移。为解决这个问题,本文结合可变形卷积组成两个对齐模块,每个模块都会选定一张图像作为参考帧,来对齐除参考帧之外的两张图像。通过两个对齐模块交叉对齐,既可以保持对齐的稳定性,又可以降低信息的冗余性。在融合模块中,结合注意力分数和可变形卷积融合来自多个图像的对齐特征(两个对齐模块的输出)。重建模块是U-Net网络,用于解决将融合模块的输出重建成高分辨率图像的问题,重建模块将融合模块的输出作为网络编码器的一部分,并将网络的细节修改成更适合图像超分辨。在模型后期,通过融合边缘图来生成大尺寸的边缘图,一方面可以约束重建网络的职能,另一方面还可以使得输出图片保持一定的纹理性。最后,经过多次实验并对比其他经典的网络模型,我们的网络模型在PSNR、SSIM和MAE等评价指标上表现良好,证明了模型的合理性。更多还原

【Abstract】 Iris recognition technology is a biometric recognition technology based on the iris in the eye for identification.Compared with other biometric methods such as fingerprint recognition and face recognition,it has higher recognition accuracy and lower misrecognition rate.Besides,iris recognition requires no repeated registration and direct contact.More importantly,it is very difficult to counterfeit an iris.All the advantages mentioned above make iris recognition ”the most reliable biometric technology” other than DNA.Compared with other biometric technologies,the biggest difficulty in the application of iris recognition technology is to collect a user’s clear binocular iris image from a distance without the need for a high degree of cooperation from the user.Compared with short-distance iris acquisition,long-distance iris acquisition urgently needs to solve a number of key technical problems from hardware to software.The diameter of the iris of the human eye is 11 mm.It is difficult for ordinary photographing equipment to acquire clear iris texture images at a distance of about 2 meters.Only special imaging systems and light source systems can solve this problem.At the same time,it is necessary to solve the iris image problems caused by long-distance iris acquisition,such as outof-focus images caused by inaccurate focus during shooting,low-resolution iris images,motion blur caused by the instability of the subject’s body,eyelids and eyelashes blocking the iris These problems are difficult to realize long-distance iris recognition.In the collected iris images,low-quality iris images account for a large proportion.Obtaining high-resolution and clear iris images from a large number of collected low-quality iris images(low resolution,out-of-focus blur,and motion blur)is a prerequisite for accurate iris identification,which requires the design of accurate and efficient iris image super-resolution reconstruction and blurred iris image restoration algorithms.Image Super-Resolution refers to the reconstruction of high-resolution images from lowresolution images.Among them,the research on Single Image Super-Resolution(SISR)is relatively extensive,and many achievements have been made.Multi-Frame Super-Resolution(MFSR)is relatively difficult due to the inherent difficulty of the problem itself.Since the large-scale application of deep learning in the field of image processing,the quality of image super-resolution has been qualitatively improved.In the field of long-distance iris recognition,due to the limitations of optical principles and image acquisition equipment,only low-resolution images can be collected in many cases,and the image quality has a great impact on the performance of the recognition model.These motivate us to study the problem of super-resolution reconstruction of iris images using deep learning methods.We collected 4300 iris images with different blur degrees,different resolutions,and different occlusion degrees through a specific long-distance iris image acquisition device as training and testing datasets.On the basis of previous research,combined with the conditions and application fields of long-distance iris acquisition,this paper constructs the overall network framework of Alignment–Fusion–Restruction.The input of multi-frame super-resolution is multiple noisy,low-resolution images with relative displacement and disorder,and the images have unknown displacement.To solve this problem,this paper combines deformable convolution to form two alignment modules,each of which selects an image as a reference frame to align two images other than the reference frame.By cross-aligning two alignment modules,the stability of alignment can be maintained and the redundancy of information can be reduced.In the fusion module,the alignment features from multiple images(outputs of two alignment modules)are fused together with attention scores and deformable convolution.The reconstruction module is a U-Net network used to solve the problem of reconstructing the output of the fusion module into a high-resolution image.The reconstruction module takes the output of the fusion module as part of the network encoder and modifies the details of the network to be more suitable for image super resolution.In the later stage of the model,the large-size edge map is generated by fusing the edge map,which can constrain the function of the reconstruction network on the one hand,and keep the output image with a certain texture on the other hand.Finally,after many experiments and comparing with other classic network models,our network model performs well on evaluation indicators such as PSNR,SSIM and MAE,which proves the rationality of the model.更多还原