近年来，互联网和智能系统技术的快速发展进一步提升了生物识别的市场需求，对生物识别系统的准确率和便捷度等性能提出了更加严格的要求。掌纹中包含丰富的多种类别生物特征，在特征量、独特性、稳定性、隐私性等方面具有优势，掌纹识别是一种有广阔应用前景的生物识别技术。然而，目前国内外已有的研究在低质量识别、跨模态识别方面存在不足，无法满足掌纹识别系统广泛部署应用的需要。为此，论文提出了基于生成式对抗网络的多模态掌纹建模研究框架。在全面综述和分析国内外掌纹识别及图像修复方法的基础上，论文将任务划分为两个类别：单模态建模和跨模态转换建模。在此基础上，论文聚焦二维低分辨率漫反射掌纹、二维高分辨率漫反射掌纹、二维高分辨率全反射掌纹和三维高精度点云掌纹四种模态的图像，针对含噪声掌纹图像识别、采集掌纹图像缺失、小面积重叠掌纹识别和二维掌纹跨模态识别的具体问题，提出了相适应的建模方法。 在单模态建模研究方面，论文通过对高质量掌纹图像添加噪声、裁剪等方法以模拟低质量掌纹图像中存在的噪声和缺损，从而构建单模态掌纹去噪和修复数据集。之后，论文使用该数据集训练深度卷积掌纹去噪、修复单向或双向网络，并借助对抗损失、Gabor损失、多隐空间感知损失等，确保语义推理的结果符合真实掌纹样本的分布。去噪和修复操作使得识别算法在处理含噪声、缺损和小面积重叠的掌纹图像时的表现明显提升，对提升单模态掌纹识别鲁棒性具有重要意义。 在跨模态转换建模研究方面，论文通过设计接触式或非接触式双模态掌纹采集设备，同时采集两种模态的掌纹图像，然后使用直接或间接的方式构建具有像素级对应关系的目标双模态掌纹跨模态转换数据集。之后，论文使用该数据集训练目标双模态掌纹图像之间的跨模态转换深度卷积网络，在保留共享结构信息的同时实现图像风格的迁移。跨模态掌纹转换操作使得掌纹识别系统可以结合多种模态掌纹图像的优点，为提升系统便捷性、兼容性以及实现跨模态识别奠定基础。 在以上场景的多模态掌纹图像处理与识别研究，突出了论文的两个方面的创新点：一是优化了单模态掌纹识别，增强系统处理低质量图像的鲁棒性；二是突破了跨模态掌纹识别，首次建立多模态掌纹之间的联系，整合多模态掌纹的优势。实验结果分析表明，论文提出的多模态建模方法有效处理了现有掌纹识别系统中的对应问题，为多模态掌纹识别系统的部署应用提供新的技术保障和发展思路。

In recent years, the rapid development of the Internet and intelligent system technology has further increased the demand for biometrics recognition. At the same time, more stringent requirements have been put forward for the accuracy, convenience and security of the biometric system. Palmprint contains a variety of features, and has advantages in feature quantity, uniqueness, stability, privacy, etc. Therefore, palmprint recognition is a biometric technology with broad application prospects. However, the existing researches at home and abroad have shortcomings in low-quality recognition and cross-modal recognition, and cannot meet the needs of the wide range application of palmprint recognition systems. To address these problems, this dissertation constructs a multi-modal palmprint modeling framework based on generative adversarial networks.Based on a comprehensive review and analysis of palmprint recognition and image inpainting methods at home and abroad, the dissertation divides the multimodal palmprint modeling problem into two parts, namely single-modal modeling and cross-model translation modeling. On this basis, the dissertation focuses on the palmprint of four modalities: 2D low-resolution diffuse reflection palmprint, 2D high-resolution diffuse reflection palmprint, 2D high-resolution total internal reflection palmprint, 3D high-precision point cloud palmprint.To solve some specific problems in multimodal palmprint recognition, such as noisy palmprint images, incomplete palmprint acquisition, small-area overlapping palmprint and cross-modal palmprint, the dissertation proposes corresponding generative methods. In terms of single-modal modeling, the dissertation simulates denoising and restoration databases by adding noise or cropping the images in existing high-quality palmprint databases. After that, these databases are used to train the deep convolutional networks. Besides, adversarial loss, Gabor loss and multiple hidden space perceptual loss are designed to ensure that the results of semantic inference comform to the distribution of real palmprint samples. Experimental results show that the denoising and inpainting operations can significantly imporve the performance of the recognition algorithms when dealing with palmprint iamges with noise or small-area overlapping, which is beneficial to imporve the robustness of single-modal palmprint recognition. In terms of cross-modal translation modeling, the dissertation collects dual-modal palmprint images using designed acquisition device, and then directly or indirectly constructs coss-modal translation databases with pixel-level correspondence. After that, these databases are used to train the cross-modal translation networks, which can achieve image style transfer while preserving shared structural information. The cross-modal palmprint translation operation enables the palmprint recognition system to combine the advantages of multimodal palmprint, laying a foundation for improving system convenience, compatibility and realizing cross-modal recognition. The multi-modal palmprint processing and recognition research in the above scenarios shows two innovations of this dissertation: first, it improves the performance of single-modal palmprint recognition and enhances the robustness of the system when dealing with low-quality images; second, it realizes cross-modal palmprint recognition, establishes the connection between multimodal palmprints and integrates the advantages of multimodal palmprints for the first time. The relevant experimental results show that the multimodal modeling method proposed in this dissertation can effectively deal with the corresponding problems in the existing palmprint recognition system, and provide some new technical guarantee and strategies for the application of multimodal palmprint recognition systems.