随着光电成像技术的发展和计算机算力的提升，成像系统的核心架构正逐步由前端硬件设备向后端计算重构技术转移，形成了计算光学成像领域。计算光学成像不同于传统相机的物像关系和结构形态，设计与功能灵活多样，可实现多维复杂光场感知。无透镜成像技术作为计算光学成像的一个重要分支，采用轻薄化光学元件代替透镜或透镜组对场景图像进行编码，可以显著减小成像系统的体积和重量，在嵌入式系统、可穿戴设备、内窥镜等领域有着广阔的应用前景。当前，无透镜成像技术在重建质量和计算效率上还无法满足实际应用需求，而深度学习技术的发展有望突破无透镜成像技术的瓶颈。本论文围绕非相干无透镜成像技术中图像重建质量和计算效率提升两个核心问题，开展了基于深度学习的菲涅尔孔径编码和无透镜光纤束成像技术研究。 提出了单帧菲涅尔孔径编码成像无孪生像重建方法，分析了重建图像中孪生像的生成机制，构建了全变差正则化下的菲涅尔孔径编码成像重建模型，通过迭代优化算法消除了孪生像噪声。构建了无需校准的无透镜相机样机，在实验中实现了对二值、灰度和彩色图像重建质量的提升。进一步建立了基于部分采样的编码掩膜成像模型，设计了基于交替方向乘子法的压缩重建算法，仅使用少量测量数据即可恢复清晰度良好的图像。 提出了基于深度学习的菲涅尔孔径编码成像方法，针对衍射效应导致图像重建模糊问题，提出宽带光源照明下编码掩膜成像系统的点扩散函数计算方法，并将点扩散函数用于深度学习数据集的生成，避免了繁琐冗长的数据集采集流程。基于U-Net和图像超分辨网络设计了端到端的网络模型，实现了图像快速高质量重建。针对二值、灰度和彩色图像，在相同的图像重建质量下，计算速度比迭代优化算法提高了两个数量级。 提出了基于深度学习的无透镜光纤内窥镜成像方法，研究了无透镜光纤束成像模型，分析了工作距离对成像质量的影响。设计并构建了光纤束图像采集装置，实现了真值图像与光纤束图像训练图像的获取。基于U-Net和图像超分辨网络设计了光纤束成像分辨率增强网络模型，消除了光纤束图像中蜂窝状伪影，实现了单帧光纤束图像的快速高分辨成像。针对肿瘤图像识别应用，提出的分辨率增强方法有效提高了胶质母细胞瘤的识别率。

With the development of photoelectric imaging technology and the improvement of computing power, the core architecture of imaging system is gradually shifting from the front-end hardware equipment to the back-end computing reconstruction technology, forming the field of computational optical imaging. Computational optical imaging is different from the object-image relationship and structure form of traditional camera, and its design and function are flexible and diverse, which can realize multi-dimensional and complex light field perception. As an important branch of computational optical imaging, lensless imaging technology uses thin optical elements instead of lenses to encode scene images, which can significantly reduce the volume and weight of the imaging system, and has broad application prospects in embedded system, wearable devices, endomicroscopy, etc. At present, lensless imaging technology cannot meet the requirements of practical application in terms of imaging quality and imaging speed, and the development of deep learning technology is expected to break through the bottleneck of lensless imaging technology. Focusing on two core issues of image reconstruction quality and computational efficiency in incoherent lensless imaging technology, this dissertation carried out research on Fresnel zone aperture imaging and lensless fiber bundle imaging technology based on deep learning. A twin-image-free reconstruction method for single-shot Fresnel zone aperture imaging is proposed. The generation mechanism of twin image in reconstructed images is analyzed. The Fresnel zone aperture imaging reconstruction model under total variation regularization is constructed, and the twin image noise in the back-propagation reconstruction is eliminated by an iterative optimization algorithm. A lensless camera prototype without calibration is constructed, and the quality of binary, gray and color image reconstruction is improved in the experiment. Furthermore, the encoding mask imaging model based on partial sampling is established, and the compressive reconstruction algorithm based on alternating direction method of multipliers is designed. Only a small amount of measured data is used to restore the image with good definition. A deep learning-based Fresnel zone aperture imaging method is proposed. To address the obscure image reconstruction caused by diffraction effect, a calculation method of point spread function (PSF) of encoding mask imaging system under broadband illumination is proposed. The PSF is used to generate training dataset, which avoids the tedious and lengthy dataset collection process. An end-to-end network model is designed based on U-Net and image super-resolution network to realize fast and high-quality image reconstruction. For binary, gray and color images, the computation speed is improved by two orders of magnitude compared with the iterative optimization algorithm under the same image reconstruction quality. A lensless fiber endoscopic imaging method based on deep learning is proposed. The imaging model of lensless optical fiber bundle is studied, and the effect of working distance on imaging quality is analyzed. A fiber bundle image acquisition setup is built to obtain fiber bundle images with the corresponding ground truth images for training. The resolution enhancement network model of fiber bundle imaging is designed based on U-Net and image super-resolution network, which eliminates the honeycomb artifacts in fiber bundle image and realizes the fast and high-resolution imaging of single fiber bundle image. For the application of tumor image recognition, the proposed resolution enhancement method can effectively improve the recognition rate of glioblastoma.