基于深度学习的单像素同时成像、定位和分类

Single-pixel imaging (SPI) uses a single-pixel detector instead of an array detector to obtain information about an object. SPI has aroused widespread research interest. Deep learning has powerful feature extraction capabilities. Many schemes based on deep learning have made breakthroughs in many fields of computer vision, such as image classification, object detection, and image segmentation. This project proposes a scheme named simultaneously single-pixel imaging, locating, and classification based on deep learning, which is planned to be experimentally verified. It is expected that objects with different numbers, different sizes, and overlapping can be imaged, located, and classified at the same time. The applicant intends to adopt multi-task learning technology based on feature sharing, construct a new loss function, and make a data set suitable for this project. Also, the fine-tuning technology will be used to improve the performance of deep learning networks. Based on the above research plan, this project can simultaneously complete single-pixel high-quality imaging, precisely locating, and classifying with high-accuracy under a low sampling rate. The results of this project have broad application prospects in areas where array detectors are unavailable or expensive, such as terahertz imaging, infrared imaging, and X-ray imaging, as well as remote sensing, medical treatment, security, automatic driving, and other fields.

单像素成像只需要一个单像素探测器而不需要阵列探测器就能获取物体信息，近年来引起了广泛的研究兴趣。深度学习因为强大的特征提取能力帮助计算机视觉的多个领域，比如，图像分类、目标检测、图像分割等取得了突破性进展。申请人提出基于深度学习的单像素同时成像、定位和分类的方案，拟进行实验验证，预期可实现对于不同数量、不同大小、而且有重叠的物体的同时成像、分类和定位。申请人拟采取以特征复用技术为基础的多任务学习技术，并构造新的损失函数，制作适合本项目的数据集。另外，申请人拟采用网络微调（fine-tuning）技术，提升深度学习网络的性能。基于以上研究方案，本项目可在低采样率下同时完成高质量单像素成像、高精度定位以及高准确度分类。在阵列探测器难以获得的领域，比如X光、红外光以及THz波段，以及遥感探测、医疗、安防、自动驾驶等领域，本项目成果具有广泛的应用前景。

单像素成像只需要一个单像素探测器而不需要阵列探测器就能获取物体信息。鬼成像，也称为关联成像、量子成像，是一种利用光的高阶关联效应获取物体信息的技术。计算鬼成像不需要参考光，与单像素成像具有极为密切的联系。深度学习因为强大的特征提取能力帮助计算机视觉的多个领域，比如图像分类、目标检测、图像分割等取得了突破性进展。本项目基于特征复用技术为基础的多任务深度学习技术，构造了新的损失函数，制作了适合本项目的数据集，实现了低采样率条件下，对于不同数量、不同大小、而且有重叠的物体的同时高质量成像、高精度定位、和高准确率分类，预测速度可以达到视频级（大于25Hz）。在本项目的支持下，还提出了一种关联成像新机制——导数鬼成像，首次完成了光强导数高阶关联特性的研究和成像能力的验证，在无需压缩感知和机器学习等先进算法的前提下，可将成像信噪比提升一倍。速度像在自动驾驶和遥感探测等领域具有重要应用，在本项目的支持下，首次提出了单像素同时成全场速度像和空间像的方法，实验实现了测速范围2.66 cm/s 到1327.34 cm/s，测速分辨率2.66 cm/s的单像素同时测速成像系统。