Handwritten digit recognition based on ghost imaging with deep learning

We present a ghost handwritten digit recognition method for the unknown handwritten digits based on ghost imaging(GI) with deep neural network, where a few detection signals from the bucket detector, generated by the cosine transform speckle, are used as the characteristic information and the input of the designed deep neural network(DNN), and the output of the DNN is the classification. The results show that the proposed scheme has a higher recognition accuracy(as high as98% for the simulations, and 91% for the experiments) with a smaller sampling ratio(say 12.76%). With the increase of the sampling ratio, the recognition accuracy is enhanced. Compared with the traditional recognition scheme using the same DNN structure, the proposed scheme has slightly better performance with a lower complexity and non-locality property.The proposed scheme provides a promising way for remote sensing.     

Computational ghost imaging with deep compressed sensing

Computational ghost imaging (CGI) provides an elegant framework for indirect imaging, but its application has been restricted by low imaging performance. Herein, we propose a novel approach that significantly improves the imaging performance of CGI. In this scheme, we optimize the conventional CGI data processing algorithm by using a novel compressed sensing (CS) algorithm based on a deep convolution generative adversarial network (DCGAN). CS is used to process the data output by a conventional CGI device. The processed data are trained by a DCGAN to reconstruct the image. Qualitative and quantitative results show that this method significantly improves the quality of reconstructed images by jointly training a generator and the optimization process for reconstruction via meta-learning. Moreover, the background noise can be eliminated well by this method.     

Efficient implementation of x-ray ghost imaging based on a modified compressive sensing algorithm

Towards efficient implementation of x-ray ghost imaging (XGI), efficient data acquisition and fast image reconstruction together with high image quality are preferred. In view of radiation dose resulted from the incident x-rays, fewer measurements with sufficient signal-to-noise ratio (SNR) are always anticipated. Available methods based on linear and compressive sensing algorithms cannot meet all the requirements simultaneously. In this paper, a method based on a modified compressive sensing algorithm with conjugate gradient descent method (CGDGI) is developed to solve the problems encountered in available XGI methods. Simulation and experiments demonstrate the practicability of CGDGI-based method for the efficient implementation of XGI. The image reconstruction time of sub-second implicates that the proposed method has the potential for real-time XGI.     

基于阈值分割的并行压缩鬼成像方法

为了提高鬼成像的速度,提出了基于阈值分割的并行压缩差分鬼成像方法。首先通过设置合理的阈值筛选出涨落较为明显的数据进行采样;然后再通过并行分块降低图像维度,提高整体运算速度;最后通过压缩感知重构算法重构图像。通过对图"BUAA"的半物理仿真以及透射物体"光"的实验结果表明,与普通的二阶赝热光关联成像相比,该方法能以更少的采样次数和运行时间重构图像,可提升成像信噪比。     

Optical encryption scheme based on ghost imaging with disordered speckles

An optical encryption scheme based on a ghost imaging system with disordered speckles is proposed to obtain a higher security with a small key. In the scheme, Alice produces the random speckle patterns and obtains the detection results with the help of a computational ghost imaging(CGI) system. Then Alice permutes the order of the random speckle patterns and shares the permutation sequence as a secure key to the authorized users. With the secure key, Bob could recover the object with the principle of the CGI system, whereas, the unauthorized users could not obtain any information of the object. The numerical simulations and experimental results show that the proposed scheme is feasible with a small key, simultaneously,it has a higher security. When the eavesdropping ratio(ER) is less than 40%, the eavesdropper cannot acquire any useful information. Meanwhile, the authorized users could recover completely with the secure key.     

Reconstruction method of computational ghost imaging based on non-local generalized total variation北大核心CSCD

鬼成像是一种能够透过大雾等恶劣环境的成像技术。针对传统鬼成像重建图像存在噪声较多、图像对比度较低等问题,将非局部广义全变分方法用于鬼成像的图像重建之中,提出基于非局部广义全变分的计算鬼成像重建方法。所提方法构造了一种非局部相关性权重设计梯度算子,将其代入全变分重建算法中,使得重建的图像能有效去除噪声的同时实现细节较好的还原。首先在不同条件下进行仿真模拟,得到所提方法的峰值信噪比相对其他方法提升1 dB左右,且具有更好的主观视觉效果,进而设计并搭建实验平台对算法的有效性进行验证,实验结果证明了所提方法在去除噪声和细节重建等方面的优越性。     

基于自适应阈值方法实现迭代降噪鬼成像

为了有效降低传统鬼成像中相关噪声对成像质量的影响,本文提出一种基于最佳阈值的迭代降噪鬼成像.首先在迭代降噪鬼成像的基础上,采用自适应阈值迭代法,在不需要目标先验信息的前提下,找到一个逼近传统鬼成像中相关噪声的阈值,根据得到的阈值构造噪声干扰项.为了每次迭代初值更接近原始目标的透射系数,对其进行二值化,降低重构图像背景噪声对迭代性能的影响.仿真以及实验结果表明,本文提出的方法与传统方法相比,视觉效果以及峰值信噪比值有明显提高.     

Correspondence normalized ghost imaging on compressive sensing

Ghost imaging(GI) offers great potential with respect to conventional imaging techniques. It is an open problem in GI systems that a long acquisition time is be required for reconstructing images with good visibility and signal-to-noise ratios(SNRs). In this paper, we propose a new scheme to get good performance with a shorter construction time. We call it correspondence normalized ghost imaging based on compressive sensing(CCNGI). In the scheme, we enhance the signal-to-noise performance by normalizing the reference beam intensity to eliminate the noise caused by laser power fluctuations, and reduce the reconstruction time by using both compressive sensing(CS) and time-correspondence imaging(CI) techniques. It is shown that the qualities of the images have been improved and the reconstruction time has been reduced using CCNGI scheme. For the two-grayscale "double-slit" image, the mean square error(MSE) by GI and the normalized GI(NGI) schemes with the measurement number of 5000 are 0.237 and 0.164, respectively, and that is 0.021by CCNGI scheme with 2500 measurements. For the eight-grayscale "lena" object, the peak signal-to-noise rates(PSNRs)are 10.506 and 13.098, respectively using GI and NGI schemes while the value turns to 16.198 using CCNGI scheme. The results also show that a high-fidelity GI reconstruction has been achieved using only 44% of the number of measurements corresponding to the Nyquist limit for the two-grayscale "double-slit" object. The qualities of the reconstructed images using CCNGI are almost the same as those from GI via sparsity constraints(GISC) with a shorter reconstruction time.     

基于NPU的实时深度学习跟踪算法实现北大核心CSCD

深度学习在检测领域高速发展,但受限于训练数据和计算效率,在基于嵌入式平台的边缘计算领域,尤其是实时跟踪应用中深度学习的智能化算法应用并不广泛。针对这一现象,同时为满足现阶段国产化、智能化的技术需求,提出了一种改进的孪生网络深度学习跟踪算法。在特征网络加入微调网络,解决了网络模型无法在线更新的问题,提升了跟踪的准确性;在IoUNet损失函数中加入中心距离惩罚项,解决了IoUNet当IoU相同时位置跳跃,存在收敛盲区和收敛速度慢的问题;将训练后的网络通过通道剪枝,缩减网络模型尺寸,提升了模型加载和运行的速度。在华为Atlas200NPU平台上实现了实时运行,算法准确率高达0.90(IoU>0.7),帧率达到66 Hz。     

基于深紫外激光-光发射电子显微技术的高分辨率磁畴成像

基于磁二色效应的光发射电子显微镜磁成像技术是研究薄膜磁畴结构的一种重要研究手段,具有空间分辨率高、可实时成像以及对表面信息敏感等优点.以全固态深紫外激光(波长为177.3 nm;能量为7.0 eV)为激发光源的光发射电子显微技术相比于传统的光发射电子显微镜磁成像技术(以同步辐射光源或汞灯为激发源),摆脱了大型同步辐射光源的限制;同时又解决了当前阈激发研究中由于激发光源能量低难以实现光电子直接激发的技术难题,在实验室条件下实现了高分辨磁成像.本文首先对最新搭建的深紫外激光-光发射电子显微镜系统做了简单介绍.然后结合超高真空分子束外延薄膜沉积技术,成功实现了L10-FePt垂直磁各向异性薄膜的磁畴观测,其空间分辨率高达43.2 nm,与利用X射线作为激发源的光发射电子显微镜磁成像技术处于同一量级,为后续开展高分辨磁成像提供了便利.最后,重点介绍了在该磁成像技术方面取得的一些最新研究成果:通过引入Cr的纳米"台阶",成功设计出FePt的(001)与(111)双取向外延薄膜;并在"台阶"区域使用线偏振态深紫外激光观测到了磁线二色衬度,其强度为圆二色衬度的4.6倍.上述研究结果表明:深紫外激光-光发射电子显微镜磁成像技术在磁性薄膜/多层膜体系磁畴观测方面具备了出色的分辨能力,通过超高真空系统与分子束外延薄膜制备系统相连接,可以实现高质量单晶外延薄膜制备、超高真空原位传输和高分辨磁畴成像三位一体的功能,为未来磁性薄膜材料的研究提供了重要手段.     

Resolution and noise in ghost imaging with classical thermal light

The resolution and classical noise in ghost imaging with a classical thermal light are investigated theoretically. For ghost imaging with a Gaussian Schell model source, the dependences of the resolution and noise on the spatial coherence of the source and the aperture in the imaging system are discussed and demonstrated by using numerical simulations. The results show that an incoherent source and a large aperture will lead to a good image quality and small noise.     

Secret shared multiple-image encryption based on row scanning compressive ghost imaging and phase retrieval in the Fresnel domain

A multiple-image encryption method is proposed that is based on row scanning compressive ghost imaging, (t, n) threshold secret sharing, and phase retrieval in the Fresnel domain. In the encryption process, after wavelet transform and Arnold transform of the target image, the ciphertext matrix can be first detected using a bucket detector. Based on a (t, n) threshold secret sharing algorithm, the measurement key used in the row scanning compressive ghost imaging can be decomposed and shared into two pairs of sub-keys, which are then reconstructed using two phase-only mask (POM) keys with fixed pixel values, placed in the input plane and transform plane 2 of the phase retrieval scheme, respectively; and the other POM key in the transform plane 1 can be generated and updated by the iterative encoding of each plaintext image. In each iteration, the target image acts as the input amplitude constraint in the input plane. During decryption, each plaintext image possessing all the correct keys can be successfully decrypted by measurement key regeneration, compression algorithm reconstruction, inverse wavelet transformation, and Fresnel transformation. Theoretical analysis and numerical simulations both verify the feasibility of the proposed method.     

Resolution and visibility of two-color pseudothermal lensless ghost imaging

Two-color (or nondegenerate-wavelength) lensless ghost imaging using pseudothermal light source is investigated theoretically by use of classical optical coherence theory. We find that for two-color pseudothermal lensless ghost imaging the visibility and resolution is determined by (Δ) the product of the wavelength and the corresponding path length rather than the wavelength for each path or the path length. We also confirm our theoretical conclusion by numerical simulations. The result shows we must make a trade-off between resolution and visibility in devise practical experiments, as the resolution of the imaging can be improved by decreasing Δ but with visibility getting poor, and vice versa.     

Lens ghost imaging in a non-Kolmogorov slant turbulence atmosphere

Based on the method of computation reference channel and the extended Huygens–Fresnel integral, the model of lens ghost imaging with fully spatially incoherent linear polarization light through a slant turbulent channel has derived. The model shows that the resolution ratio of ghost imaging decreases as the power-law exponent of non-Kolmogorov turbulence increasing or the object location departing from the source. The zenith angle of channel has little influence to the quality of the ghost imaging. The minimum distinguishable centre-separation of double slit decreases as the power-law exponent of non-Kolmogorov turbulence increasing.     

常压空气亚微秒脉冲DBD高速摄影研究

为研究亚微秒脉冲激励常压空气介质阻挡放电（DBD）时空演化规律,采用高速摄影相机(ICCD)在ns量级时间曝光范围内,分别在单次及连续放电（500 Hz）情况下研究放电等离子体的发光性质及放电发展过程。结果表明：在亚微秒高压脉冲平行板结构放电条件下,起始阶段放电表现出不均匀性,随着电流增大发光增强并变得均匀,在回路电流最强位置时放电达到均匀状态,并且放电均匀性在回路电流下降阶段明显好于其上升沿阶段,放电发光强度与电流波形可以很好的吻合。在一定外界参数条件下,二次放电的存在受到电源重复频率、电压幅值等参数的影响。在单次放电条件下可以明显观测到二次放电的存在,相较于一次放电,二次放电更加均匀;连续状态放电模式下几乎观察不到二次放电。     

常压空气亚微秒脉冲DBD高速摄影研究

为研究亚微秒脉冲激励常压空气介质阻挡放电(DBD)时空演化规律,采用高速摄影相机(ICCD)在ns量级时间曝光范围内,分别在单次及连续放电(500Hz)情况下研究放电等离子体的发光性质及放电发展过程。结果表明:在亚微秒高压脉冲平行板结构放电条件下,起始阶段放电表现出不均匀性,随着电流增大发光增强并变得均匀,在回路电流最强位置时放电达到均匀状态,并且放电均匀性在回路电流下降阶段明显好于其上升沿阶段,放电发光强度与电流波形可以很好的吻合。在一定外界参数条件下,二次放电的存在受到电源重复频率、电压幅值等参数的影响。在单次放电条件下可以明显观测到二次放电的存在,相较于一次放电,二次放电更加均匀;连续状态放电模式下几乎观察不到二次放电。     

Imaging a periodic moving/state-changed object with Hadamard-based computational ghost imaging

We propose a method for imaging a periodic moving/state-changed object based on computational ghost imaging with Hadamard speckle patterns and a slow bucket detector, named as PO-HCGI. In the scheme, speckle patterns are produced from a part of each row of a Hadamard matrix. Then, in each cycle, multiple speckle patterns are projected onto the periodic moving/state-changed object, and a bucket detector with a slow sampling rate records the total intensities reflected from the object as one measurement. With a series of measurements, the frames of the moving/state-changed object can be obtained directly by the second-order correlation function based on the Hadamard matrix and the corresponding bucket detector measurement results. The experimental and simulation results demonstrate the validity of the PO-HCGI. To the best of our knowledge, PO-HCGI is the first scheme that can image a fast periodic moving/state-changed object by computational ghost imaging with a slow bucket detector.     

Full color ghost imaging by using both time and code division multiplexing technologies

We propose a new full color ghost imaging scheme using both time and code division multiplexing technologies. In the scheme, the speckle patterns of three colors (red, green and blue) are modulated with different time slots and codes. The light intensity is sampled by one bucket detector. Then based on the modulated time slots and codes, we can effectively and simultaneously extract three detection component signals corresponding to three color components of objects from the sampling signal of the bucket detector. Finally, three component images resulting from the three component detection signals can be synthesized into a full color image. The experimental results verify the feasibility of our scheme under the limit of the number of time slots and codes. Moreover, our scheme reduces the number of bucket detectors and can realize high quality imaging even in a noisy environment.     

三维混合域重建算法在太赫兹全息成像中的应用

提出了一种三维混合域重建算法,该算法在垂直方位向采用空间波数域匹配滤波,结合距离水平方位域后向投影技术完成目标三维重建。与三维波数域算法相比,该算法不仅避免了复杂的三维Stolt插值,还继承了后向投影技术便于处理运动补偿的优点。相比传统的三维时域后向投影算法,该算法通过在垂直方位向的波数域处理,运算量显著减少。通过对0.2 THz波段仿真和实验数据的重建,验证了算法的正确性和有效性。     

基于多域特征提取和深度学习的声源被动测距*

由于实际海洋环境中存在大量的非高斯噪声,一些基于高斯假设的传统去噪方法在实际海洋环境中性能下降甚至失效。针对非高斯噪声,如α稳定分布噪声、非平稳行船噪声下的脉冲信号的去噪与重构,该文提出一种基于深度学习的方法。去噪模型首先通过学习带噪信号短时傅里叶变换谱与残差谱之间的映射关系以去除环境噪声,之后对去噪信号的时频谱进行逆变换重构脉冲信号。仿真实验结果表明,深度学习模型在非高斯噪声环境下脉冲信号的去噪与重构任务中有着良好的表现,在实测样本上也表现出良好的泛化性,体现了一定的工程应用价值。