基于深度学习的移位MIMO“鬼"成像方法研究

Ghost imaging (GI), also named correlated imaging, is currently a hot topic in quantum optics. Based on the correlation of entanglement or field fluctuations, one can obtain a clear image of a non-localized object, at the idler path, by the intensity correlation between the signal beam and idle beam. GI has huge values and extensive usages in optically harsh (difficult to be reached) or noisy environments. However, it needs take thousands measurement times to get the data from the bucket detector and cost a lot of time to recover the imaging in current GI system. In the proposal, we will study on novel ghost imaging schemes by taking advantage of multiple-input-multiple-output (MIMO), edge detection and deep learning technique, these advanced signal processing methods in information processing area. Based on the theory of MIMO radar imaging, a MIMO ghost imaging scheme is first proposed to greatly decrease the times to get the data by using ‘space’ as a substitute for ‘time’, where mutiple sources (the multiple input) are used at the same time to illustrate the unkown object, and mutiple bucket detectors (the multiple output) are adopted simultaneously to get the data. With the prior information of the object, the effective edge detection methods in space domain and frequency domain are designed and applied to the MIMO ghost imaging scheme to directly obtain the edge information of the unknown object. A deep neural network is trained to solve the end-to-end inverse problems in ghost imaging, and a speckle-shifting MIMO ghost imaging scheme using deep learning is presented. At last, the experiments are implemented to testify the proposed schemes. With MIMO, edge detection and deep learning techniques, our proposed schemes will provide more accurate edge imaging with a snapshot. The results are significant and valuable to distributed sensing, detection, and distributed image processing etc.

“鬼”成像，即关联成像，是基于纠缠或光场涨落的关联特性，通过强度关联非局域地获取目标物体信息的新型成像方法，具有巨大应用前景。然而，现有方案需要成千上万次采集桶探测器数据后，再通过重建算法，才能获得目标物体清晰的图像。本项目拟将多输入多输出(MIMO)、边缘检测和深度学习技术引入鬼成像，探索一种高质量瞬时可成像的方案，是信息技术和量子光学领域的交叉课题。借鉴MIMO雷达成像理论，以‘空间’换‘时间’，获得多散斑光场同时照射(多输入)、多个桶探测器同时接收（多输出）的鬼成像，解决需要长时间采集数据的问题；设计高性能空域和频域边缘检测方法，直接获取目标物体轮廓信息，避免低成像质量对目标识别的限制；在此基础上，利用深度神经网络，探索基于深度学习的移位MIMO鬼成像协同优化方案，并用实验实现该方法。研究结果将对分布式感知、探测、分布式图像处理等研究具有重要的理论指导意义和实际参考价值。

鬼成像是通过参考光场与目标探测光场之间的强度关联，非局域地获取目标物体信息的一种新型成像技术，如何减少采集数据的次数和提高重建算法的有效性是其重要的研究方向。本项目将多输入多输出(MIMO)、边缘检测和深度学习技术引入鬼成像，探索一种高质量瞬时可成像的方案，是信息技术和量子光学领域的交叉课题。研究中，提出了基于正交时空散斑编码的多输入单输出鬼成像和基于带有特定频率余弦波形的红、绿、蓝光场散斑的全彩色单像素成像方案，有效地减少桶探测器数量和测量次数；提出空域和频域中基于鬼成像技术的边缘检测方法，在不需要获取目标物体条件下直接获得其边缘信息；将深度学习应用于恢复算法重建，获得基于线性回归、深度神经网络和整体嵌套网络的鬼成像方案，提升了重建算法的有效性，减少了鬼成像重建时间；将信号处理技术应用于鬼成像，提出了基于预加重技术、直接序列编码技术、均值滤波，以及时分复用和码分复用的方案，获得了抗噪声、成像质量高的鬼成像方法；考虑鬼成像在图案加密中的应用，提出秘钥量少、安全性高的分数傅立叶变换鬼成像、扩频鬼成像光学加密方案和秘密共享方案；提出提升成像质量方法，给出通过QR分解、差分计算优化测量散斑和傅里叶频谱采集等实现高分辨率、高性能的鬼成像方案。在项目完成期间，培养了11名博士生 (4人毕业)和31名硕士生 (21人毕业)；在国内外权威期刊、会议上发表学术论文73篇，其中53篇被SCI收录，12篇被EI收录；提交发明专利申请8项，获授权专利6项；获软件著作权登记证书1项和江苏省自然科学奖三等奖1项。研究结果将对分布式感知、探测、分布式图像处理等研究具有重要的理论指导意义和实际参考价值。