Coherent L2S THz Imaging Systems

在电磁波谱的长波区域，如毫米波和太赫兹频率范围内，相干成像的性能不断提高，为这种传感器技术的广泛采用开辟了道路。考虑到这种波长在光学难以接近的情况下成像和传感的能力，例如任意环境条件下的遥感、地下成像无损检测、弹性自动驾驶汽车视觉系统或与安全相关的成像系统，如机场检查站的隐藏爆炸物检测，应用潜力是巨大的。在所有这些应用场景中，相干成像具有巨大的优势，可以获得光学成像系统无法获得的信息，并且从根本上能够直接从相位数据中获得三维物体和场景信息。然而，这些优势是以所有相干照明和成像系统的基本成本为代价的，这些系统必须处理干扰伪影，如环、斑点和多径干扰，这些伪影会完全抑制图像的形成，并且在该频率范围内特别强烈。因此，由于这些基本的限制，这种系统的应用吸收仍然是难以捉摸的。该项目计划使用基于人工智能的方法来学习应对相干成像系统的这些基本限制，并训练和验证其在毫米波和太赫兹频率范围内的充分性，在这些频率范围内，系统参数的可变性对图像生成过程具有特别大的自由度。沿着L2S的思路，沿着完整的系统实现和图像分析管道建立一个真正的端到端学习范式，可以预见以下目标：-学习如何使用包含网络架构的物理知识来开发自适应合成图像生成方法，从而提高图像质量并纠正与场景相关的干扰伪影，以实现连贯的3D成像。-评估和理解基于机器学习的分割重建3D太赫兹成像数据的鲁棒性，这些数据来自稀疏照明和传感器布置，包括差分成像模式。-评估和学习是否可以直接从原始感官数据中获得分割，而无需通过合成重建进行中间的3D图像生成步骤。-学习感官任务依赖系统如何适应能够最大化成像和识别能力，同时最大限度地减少硬件和数据采集工作。这些目标将通过实验解决，使用基于MIMO（多输入多输出）FMCW（调频连续波）合成图像重建方法的太赫兹成像系统，以允许系统配置的最大可变性和多照明信号可变性，并作为合作伙伴开发的L2S方法的测试场景。

The increasing performance of coherent imaging in the long wavelength region of the electromagnetic spectrum, like the mm-wave and THz frequency ranges, has opened-up the path to a wide uptake of such sensor technologies. Application potential is enormous, given the capability of such wavelengths to image and sense in optically inaccessible situations, like inter alia remote sensing in arbitrary environmental conditions, subsurface imaging non-destructive testing, resilient autonomous car vision systems, or security related imaging systems like hidden explosives detection at airport checkpoints. In all such application scenarios coherent imaging has the immense advantage to attain information not accessible by optical imaging systems, and to be fundamentally capable to derive 3D object and scene information directly from phase data. However, such advantages are attained at the fundamental cost associated with all coherent illumination and imaging systems of having to cope with interference artifacts like ringing, speckles and multi-path interferences, which can totally inhibit image formation, and which are particularly strong in this frequency range. Application uptake for such systems remains elusive, therefore, given such fundamental restrictions. This project plans to use artificial intelligence-based approaches to learn to cope with such fundamental limitations of coherent imaging systems, and to train and validate their adequacy in the mm-wave and THz frequency ranges, where the system parameter variability has a particularly large degree of freedom for the image generation process. Along the idea of L2S to establish a true end-to-end learning paradigm along the complete system realization and image analysis pipeline, following goals are foreseen:- Learning how physical knowledge containing network architectures can be used to develop adaptive synthetic image generation approaches that enhance the image quality and correct scene dependent interference artifacts for coherent 3D imaging.- Evaluating and understanding the robustness of machine-learning based segmentation of reconstructed 3D THz imaging data originating from sparse illumination and sensor arrangements, including differential imaging modes.- Assessing and learning if segmentation can be attained directly from raw sensory data, without the intermediate 3D image generation step by synthetic reconstruction. - Learning how a sensory task dependent system adaptation can maximize imaging and recognition capabilities, and at the same time minimize hardware and data acquisition effort.These goals will be experimentally addressed using THz imaging systems based on MIMO (multiple-input multiple-output) FMCW (frequency-modulated continuous-wave) synthetic image reconstruction approaches, in order to allow a maximum variability of the system configuration and multiple-illumination signal variability and as a test scenario for the L2S methodologies developed by the partners.