基于卷积神经网络加速的计算鬼成像物体识别仿真

Simulation of Computational Ghost Image Object Recognition Accelerated by Convolutional Neural Networks

摘要: 鬼成像是利用两束光的空间相关性来进行成像的一种新的成像方式，具有成像方式灵活，探测器易获取等优势。作为其中的重要分支，计算鬼成像直接采用随时间变化的散斑照明图像，利用多个散斑模式和穿过物体后的总光子计数之间的关联性来计算恢复图像，可结合单像素相机，在遥感、光学加密以及边缘检测等方向发挥重要作用，同时使得通过计算机仿真鬼成像结果成为可能。然而，为了获取高的成像质量，切换大量散斑图是必要的，这阻碍了它在快速成像和图像处理等方面的应用，因此，快速获取物体信息成为计算鬼成像的核心。基于此，本文设计了一个带有全连接层的四层卷积神经网络(CNN)用于计算鬼成像图像的快速识别，可以在一些物体识别比成像更重要的场景下(如医学图像诊断)发挥作用。通过对差分鬼成像的数值仿真结果的分析，证明了所提出的卷积神经网络可以在散斑数量是相对于完整测量的1/20的情形下，实现对低分辨率物体的接近1的识别准确度。使得CGI物体识别的等效采样率约为0.05，这大大提高了计算鬼成像中的信息获取速度。这种先于成像过程的卷积神经网络可广泛应用于与经典和量子鬼成像及其他计算成像方案如单像素成像，散射成像和无透镜成像等。

Abstract: Ghost imaging is a new paradigm of imaging that harness the spatial correlation of the two separated light fields for reconstruction of an unknown image. It has the advantages of flexible imaging strategies and easily accessible imaging equipment. As an important branch, computational ghost imaging (CGI) directly uses time-varying speckles to illuminate anobject, utilizing the correlation between multiple speckle patterns and the total photon counts after passing through the object to calculate the restored the image. It can be combined with a single pixel camera to play an important role in various areas such as remote sensing, optical encryption, and edge detection, meanwhile providing the possibility to simulate the ghost imaging process through computer calculation. However, in order to achieve a high image contrast, it is necessary to switch a large number of speckle patterns, which hinders its application in fast imaging and image processing. Therefore, quickly obtaining object information has become the core of computational ghost imaging. Therefore, quickly obtaining object information has become the core of computing ghost imaging. Here, we designed a four-layer convolutional neural network (CNN) with fully connected layers for fast recognition of ghost imaging images, which can be effective in the scenarios (medical image diagnosis for instance) that recognition of objects is more important than imaging. Through the analysis of numerical simulation results of differential ghost imaging under different speckle numbers, we proved that the proposed convolutional neural network can achieve near-unity recognition accuracy of the low-resolution objects restored with a 1/20^th of the number of speckles relative to the complete measurement. This leads to an equivalent sampling ratio of approximately 0.05 for the CGI object recognition, which greatly improves the speed of information acquisition in computing ghost imaging. This recognition method prior to the imaging process can be widely used in classical and quantum ghost imaging as well as other computation imaging schemes like single-pixel imaging, scattering imaging and lensless imaging.