基于光栅阵列和深度学习的偏振光谱成像理论研究

There is a demand for spectropolarimetric imager with compact, light, small, snapshot and high resolution used on micro-platform such as intelligent device and Micro-nano satellite. A spectropolarimetric imaging theory is proposed in this project based on grating array, and the key scientific problems are studied in this project. The guided mode resonance grating can regulate the polarization and spectrum, with narrow bandwidth, high transmittance, high extinction ratio. By utilizing the advantages of guided mode resonance grating in spectral resolution, extinction ratio and energy extraction and the advantages of deep learning in high dimensional image features extraction and representation, we propose a guided mode resonance grating array high resolution snapshot spectropolarimetric imaging technology. In the framework of rigorous coupled wave theory and the spectrum characteristics of the array, the grating array parameters are designed. And deep learning is used in mosaic spectropolarimetric image processing by analyzing the information transmission to complementary relationship in multi-task. Then multi-task integration processing method is proposed based on deep learning for spectropolarimetric mosaic image denoising, reconstruction and interpretation. The project is to realize high resolution fast acquisition of spectropolarimetric images, which lays the foundation for the establishment of new spectropolarimetric imaging system.

针对手持智能设备、微纳卫星等微小型平台对偏振光谱成像仪提出快照、紧凑、轻小、高分辨等的需求，本课题提出一种新型的偏振光谱成像方式，并对其中的关键科学问题展开研究。本课题结合导模共振光栅可以同时调节偏振和光谱，且具有带宽窄、透过率高、消光比高等优势，利用深度学习在高维图像特征提取、表示方面的极强能力，提出并研究基于导模共振光栅阵列和深度学习的偏振光谱成像。基于严格耦合波理论和滤光片阵列频谱特性，解决光栅阵列的参数设计问题；利用深度学习在马赛克偏振光谱图像处理上的优势，分析多任务之间的信息传递和互补促进关系，研究基于深度学习的马赛克偏振光谱图像去噪、重构和解译的多任务统合处理等问题，并在此基础上对所重构的偏振光谱图像进行质量评价，进而实现高分辨快照式偏振光谱图像的获取，为建立新型偏振光谱成像体制奠定基础。

针对手持智能设备、微纳卫星等微小型平台对偏振光谱成像仪提出快照、紧凑、轻小、高分辨的需求，本课题提出一种新型的偏振光谱成像方式，并对其中的关键科学问题展开研究。本课题结合导模共振光栅可以同时调节偏振和光谱，且具有带宽窄、透过率高、消光比高等优势，利用深度学习在高维图像特征提取、表示方面的极强能力，研究基于导模共振光栅阵列和深度学习的偏振光谱成像方法。主要成果如下：基于严格耦合波理论和滤光片阵列频谱特性，提出高衍射率偏振光谱光栅阵列设计方法和偏振光谱光栅阵列排布模式最优设计方法，实现了高透过、窄带宽、低串扰的导模共振光栅阵列的设计，解决光栅阵列的参数设计问题，基于所设计的光栅阵列参数研制出分焦平面偏振、光谱成像仪；利用深度学习在马赛克偏振光谱图像处理上的优势，分析多任务之间的信息传递和互补促进关系，提出偏振光谱马赛克图像多维度深度学习模型、空-谱-偏联合约束的图像分辨率重构模型和基于偏振光谱理化先验约束的深度学习模型，并基于这些模型实现了偏振光谱马赛克图像去噪去马赛克联合处理、多维度联合重构，解决了偏振光谱图像的重构问题；基于冗余线性偏振测量的自洽性，提出无参考偏振光谱图像的质量评价指标；通过对不同目标和背景的偏振光谱统计，提出基于偏振光谱视频的目标检测、跟踪、识别方法。相应理论研究成果发表在IEEE Trans. on Cybernetics、IEEE Trans. On Image Processing、ISPRS Journal of Photogrammetry and Remote Sensing、IEEE Trans. On Neural Network and Learning System等刊物，授权中国发明专利3项，部分研究成果获陕西省教育厅特等奖1项、中国产学研合作促进会二等奖1项。利用所研制的分焦平面偏振、光谱成像仪和所提出的偏振光谱视频目标检测、跟踪、识别方法，研制出智能可视化生命探测仪，获陕西省科技工作者创新创业大赛一等奖，并开展相应产业化工作。