三维太赫兹图像异构加速处理与智能化实时识别的关键技术研究

太赫兹(THz)技术因其物理指标的比较优势，在人体安检等领域受到高度关注。限于当前成像器件空间分辨率、存储芯片的访问效率、图像分类计算的复杂性等瓶颈，导致大容量、高噪声和低分辨率的三维太赫兹深度图像的恢复重建时间长、人工辅助识别的效率低等问题。因此，太赫兹图像快速重建和实时识别的方法成为国内外太赫兹领域的研究热点，是太赫兹安检系统推广应用的关键。.本项目将依托上海市现代光学系统重点实验室及自行研制的太赫兹成像系统，从四个方面开展关键技术研究:(1)研究太赫兹图像恢复重建和二值化核心算法FFT面向CPU+GPU异构计算系统的资源映射和分块计算方法；(2)研究多通道灰度图像多维特征快速聚类算法；(3)研究太赫兹图像实时分类与识别的三维生成对抗网络深度学习模型3DGAN；(4)研究异构内存页迁移策略和能效评估模型。通过对太赫兹图像处理的理论研究和模型设计，将实现三维太赫兹图像的实时分类与识别。

Terahertz (THz) technology is attracting great attentions in the field of security check due to its outstanding comparable advantages on intrinsic physical features. However, limited by the technical bottlenecks of current image spatial resolution, access speed and bandwidth of memory chips, and complexity of computing classification and recognition on Terahertz depth images, etc., there are results in the problems of taking much long time to recover and reconstruct 3D THz images, and recognizing potential objects by manual work. Therefore, it has become a global research hotspot on detecting and recognizing non-cooperative objects in 3D terahertz depth images with huge capacity, low S/N rate and weak resolution. It becomes a crucial step to promote large scale of Terahertz products and applications.. This project will, based on the research infrastructure of Shanghai Key Lab of Modern Optical Systems within University of Shanghai for Science and Technology, and self-developed THz imaging system, investigate four key technologies. Firstly, research on resource mapping and blocking computation of FFT algorithm on GPU supported heterogeneous environment while recovering and binarization processing 3D Terahertz images; Secondly, research on rapid algorithm to cluster multiple dimensional features of multi-channel gray-scale images; Thirdly, propose three-dimensional Generative Adversarial Networks (3DGAN) deep learning model to classify and recognize Terahertz images in real-time. Finally, investigate migrating strategy of memory pages over heterogeneous computer system and the corresponding evaluation model on the energy efficiency of heterogeneous system. The outcomes of this project will investigate the infrastructural model and theory on processing terahertz images, then classify and recognize the objects of 3D THz depth images in real-time by accelerating computation on heterogeneous system.