搜索

x

留言板

姓名
邮箱
手机号码
标题
留言内容
验证码

引用本文:
Citation:

张智杰, 郭龑强, 郭晓丽, 张丽, 宋铠炜, 张明江

Ultra-fast exposure enhanced imaging with SPAD arrays based on super-resolution deep learning

ZHANG Zhijie, Guo Yanqiang, Guo Xiaoli, Zhang li, Song Kaiwei, Zhang Mingjiang
Article Text (iFLYTEK Translation)
PDF
导出引用
  • 为突破单光子成像在极短曝光条件下成像质量与速率受限的问题,本文提出一种融合超分辨深度学习的单光子成像增强神经网络,实验上对单光子阵列在超快曝光时间下所获取的低信噪比图像进行增强重构,实现了在低于0.5个光子通量条件下的高保真成像.实验上实现了对微秒至毫秒曝光时间范围内512×512像素的单光子阵列高分辨成像,并对μs量级短曝光时间下的目标图像进行重构.通过构建的单光子增强深度神经网络可显著提升图像的峰值信噪比(PSNR+7.21dB)和结构相似性指数(SSIM+0.34),远高于传统超分辨方法(Bicubic)对低曝光图像超分辨增强重构(PSNR+0.49dB/SSIM+0.05),并有效补偿了超快曝光下的成像细节丢失.并在成像曝光时间5μs时,成功实现了对5.19公里处无人机512×512像素的高分辨率弱光成像,表明该方法在远距离目标成像良好的应用潜力.
    In recent years, with wide applications of high-sensitivity single-photon detectors, especially in the field of quantum imaging and optical imaging has achieved many important results, and low-light imaging technology based on single-photon level has gradually become an important branch in high-resolution imaging systems. Currently, the main single-photon detectors are single-photon counting avalanche diode (SPAD) sensors and support pixel arrays of different sizes, ranging from single-pixel detector sizes to tens of thousands of pixel SPAD arrays. The process structure of single-pixel SPAD detectors is relatively simple, and they are often used as the first choice for low-light imaging by virtue of their high sensitivity, small size, and low cost. However, single-pixel SPADs can only detect signals at a single location due to their lack of spatial resolution and cannot provide spatial information, and they are usually used in conjunction with a spatial light modulator DMD or SLM that possesses spatial resolution, to reconstruct 2D images through compressed sensing or quantum correlation. Although the single-pixel detector can provide ns-level or even ps-level temporal resolution, it is limited by the frame rate of the spatial light modulator (SLM). For example, the fastest digital micromirror device (DMD), a type of SLM, has a frame rate of 22 kHz. It means that the imaging rate of a single-pixel camera, which typically uses a SPAD and an SLM for single-photon imaging tasks, is usually limited to the order of seconds. This makes it challenging to significantly improve the imaging rate, especially when higher imaging resolutions, such as those exceeding hundreds of thousands of pixels, are required. Assuming that the imaged object is a fast-moving dynamic target, the imaging rate in the order of seconds will inevitably cause dynamic blurring, which also poses a challenge to the fast real-time performance of the single-photon imaging system.
    The SPAD array sensor retains the excellent sensitivity, low dark count rate and high temporal resolution of single-pixel SPAD sensors. Due to the improvement of the fabrication process, multiple sensors and readout circuits are fabricated on the same chip, leading to the development of spatially resolved SPAD array camera. But the integrated design of SPAD arrays with multiple pixels and circuits inevitably causes cross-crosstalk between pixels. This crosstalk can significantly affect the accuracy of the signal. Additionally, the fill factor of such array camera is typically low. Although the fill factor can be improved by methods such as 3D stacking and microlens arrays, the utilisation of the space is still to be improved in comparison with that of single-pixel SPADs. However, it is undeniable that SPAD arrays perform well in high dynamic range photon flux detection, as well as high frame rate photon counting measurements by virtue of the parallel processing of multiple detectors, and the current commercial SPAD arrays have integrated hundreds of thousands of detector pixel units, which provide excellent spatial resolution. Unfortunately, due to the manufacturing process and the challenges in various aspects, the SPAD array camera has been used in high-quantification bit deep sampling in the mode to acquire high-resolution single-photon intensity imaging. Its exposure time is limited to the ms level. It is difficult to avoid dynamic blurring during the imaging exposure time when the motion frequency of the dynamic target reaches kHz or higher. Although the quantification bit depth can be sacrificed to shorten the minimum exposure time of the array camera to the ns level, the excessively short exposure time results in the SPAD array capturing the sparse photon data that is contaminate with a large amount of noise, necessitating the development of reliable photon denoising methods. These methods are essential for effectively separating background noise from the actual signals, thereby improving the signal-to-noise ratio of the imaging system. Therefore, the real-time performance of the imaging system at the expense of quantification sampling accuracy still needs to be further optimized.
    To address the problem of limited imaging quality and rate of SPAD arrays under very short exposure time, this paper proposes a single-photon imaging enhancement deep neural network incorporating super-resolution deep learning, which can achieve high-fidelity reconstruction of low signal-to-noise single-photon images under ultra-short exposure time by constructing a single-photon image dataset with dynamic exposure time and performing adaptive training. In the experiments, it achieves the enhanced reconstruction of low-quality fan images (PSNR/SSIM, 6.54dB/0.18) under very low-light conditions with an exposure time of only 1μs and an average photon number of less than 0.5 photons (PNSR/SSIM, 13.21dB/0.34), and the images are effectively improved by +7.21dB/+0.16 for PSNR and SSIM. The passive remote enhanced reconstruction of an unmanned aerial vehicle at 5.19 km is achieved at an imaging exposure time of 5 μs, with an effective PSNR and SSIM enhancement of +4.78 dB/+0.2. This method provides a new technical solution for SPAD arrays to achieve ultra-fast-exposure high-quality imaging.
  • [1]

    [2]

    [3]

    [4]

    [5]

    [6]

    [7]

    [8]

    [9]

    [10]

    [11]

    [12]

    [13]

    [14]

    [15]

    [16]

    [17]

    [18]

    [19]

    [20]

    [21]

    [22]

    [23]

    [24]

    [25]

    [26]

    [27]

    [28]

    [29]

    [30]

    [31]

    [32]

    [33]

    [34]

    [35]

  • [1] 朱顺凯, 袁方, 胡凯, 皮涛涛, 朱熙铖, 李程. 基于卷积神经网络的双站雷达散射截面减缩超表面设计. 必威体育下载 , doi: 10.7498/aps.74.20250109
    [2] 王帅, 李希, 姚星灿. 基于增强条纹去除的超冷原子成像. 必威体育下载 , doi: 10.7498/aps.73.20240570
    [3] 黄宇航, 陈理想. 基于未训练神经网络的分数傅里叶变换成像. 必威体育下载 , doi: 10.7498/aps.73.20240050
    [4] 施岳, 欧攀, 郑明, 邰含旭, 王玉红, 段若楠, 吴坚. 基于轻量残差复合增强收敛神经网络的粒子场计算层析成像伪影噪声抑制. 必威体育下载 , doi: 10.7498/aps.73.20231902
    [5] 刘金品, 王秉中, 陈传升, 王任. 基于深度物理启发神经网络的微波波导器件逆设计方法. 必威体育下载 , doi: 10.7498/aps.72.20230031
    [6] 付亚鹏, 孙乾东, 李博艺, 他得安, 许凯亮. 基于RCA阵列三维超快超声血流成像方法仿真研究. 必威体育下载 , doi: 10.7498/aps.72.20222106
    [7] 娄月申, 郭文军. 贝叶斯深度神经网络对于核质量预测的研究. 必威体育下载 , doi: 10.7498/aps.71.20212387
    [8] 黄颖, 顾长贵, 杨会杰. 神经网络超参数优化的删除垃圾神经元策略. 必威体育下载 , doi: 10.7498/aps.71.20220436
    [9] 朱琦, 许多, 张元军, 李玉娟, 王文, 张海燕. 基于卷积神经网络的白蚀缺陷超声探测. 必威体育下载 , doi: 10.7498/aps.71.20221504
    [10] 张航, 胡月姣, 陈嘉文, 修龙汪. 程能映射下配光平移群的深度神经网络实现. 必威体育下载 , doi: 10.7498/aps.71.20220178
    [11] 向鹏程, 蔡聪波, 王杰超, 蔡淑惠, 陈忠. 基于深度神经网络的时空编码磁共振成像超分辨率重建方法. 必威体育下载 , doi: 10.7498/aps.71.20211754
    [12] 臧佳琦, 许凯亮, 韩清见, 陆起涌, 梅永丰, 他得安. 无造影剂增强的超快超声脊髓微血管成像方法. 必威体育下载 , doi: 10.7498/aps.70.20201878
    [13] 徐启伟, 王佩佩, 曾镇佳, 黄泽斌, 周新星, 刘俊敏, 李瑛, 陈书青, 范滇元. 基于深度卷积神经网络的大气湍流相位提取. 必威体育下载 , doi: 10.7498/aps.69.20190982
    [14] 郭永峰, 谭建国. 一类非线性神经网络系统的超阈值随机共振现象. 必威体育下载 , doi: 10.7498/aps.61.170502
    [15] 李海燕, 胡云安, 任建存, 朱敏, 刘亮. 非匹配不确定交叉严反馈超混沌系统神经网络反演同步. 必威体育下载 , doi: 10.7498/aps.61.140502
    [16] 冯朝文, 蔡 理, 李 芹. 基于单电子器件的细胞神经网络实现及应用研究. 必威体育下载 , doi: 10.7498/aps.57.2462
    [17] 董继扬, 张军英, 陈 忠. 自动波竞争神经网络及其在单源最短路问题中的应用. 必威体育下载 , doi: 10.7498/aps.56.5013
    [18] 周小荣, 罗晓曙, 蒋品群, 袁五届. 小世界神经网络的二次超谐波随机共振. 必威体育下载 , doi: 10.7498/aps.56.5679
    [19] 吴忠强, 谭拂晓, 王绍仙. 基于无源化的细胞神经网络超混沌系统同步. 必威体育下载 , doi: 10.7498/aps.55.1651
    [20] 蔡 理, 马西奎, 王 森. 量子细胞神经网络的超混沌特性研究. 必威体育下载 , doi: 10.7498/aps.52.3002
计量
  • 文章访问数:  166
  • PDF下载量:  4
  • 被引次数: 0
出版历程
  • 上网日期:  2025-06-11

返回文章
返回
Baidu
map