Edge quality improvement of ghost imaging based on convolutional neural network
-
摘要: 散斑移位鬼成像边缘提取方法需要对物体进行多次的采样，才能得到高质量的边缘图。为了解决散斑移位鬼成像提取物体边缘时采样次数多和时间长的问题，在鬼成像的边缘提取实验中采用了卷积神经网络。首先用Walsh散斑对未知图像进行照射，将桶探测器收集的采样信号作为图像特征信息输入到鬼成像边缘提取网络，最后通过训练好的网络直接输出探测物体的边缘信息图，并且使用非极大值抑制算法来优化卷积神经网络的输出结果。实验结果表明，对于128×128像素的重建物体，在采样次数为1600时，鬼成像边缘提取网络输出边缘图案的信噪比和结构相似指数分别比散斑移位鬼成像的输出结果提高了5倍和2倍，成功提高了低采样率下鬼成像边缘提取的质量，降低了采样的时间。采用卷积神经网络的鬼成像边缘提取方案，有利于鬼成像在物体识别、安全检查的实际应用中进行快速高质量的边缘检测。Abstract: Scatter-shift ghost imaging edge extraction methods require multiple sampling of the object to obtain a high quality edge map. To solve the problem of many samples and long time when extracting the edge of the object by scatter-shift ghost imaging, convolutional neural network is adopted to the edge extraction experiment of ghost imaging. Firstly, the unknown image is irradiated by Walsh scattering, the sampled signal collected by the barrel detector is input to the ghost imaging edge extraction network as the image feature information, finally the edge information map of the detected object is directly outputted by the trained network, and the output of the convolutional neural network is optimized by using the non-maximum value suppression algorithm. The experimental results show that for the reconstructed object of 128×128 pixels, the signal-to-noise ratio and structural similarity index of the ghost imaging edge extraction network output edge pattern are 5 times and 2 times higher than that of the scatter-shift ghost imaging respectively when the sampling number is 1600, which successfully improves the quality of the ghost imaging edge extraction under the low sampling rate and reduces the sampling time. The ghost imaging edge extraction scheme using convolutional neural network is conducive to fast and high-quality edge detection of ghost imaging in practical applications of object recognition and security inspection.
-
Key words:
- edge detection /
- ghost imaging /
- convolutional neural network /
- deep learning /
- non-maximum suppression
-
图 1 采用Unet卷积神经网络的鬼成像边缘提取方案示意图
Figure 1. Schematic diagram of the ghost imaging edge detection scheme using Unet convolutional neural network
图 2 Unet鬼成像边缘提取网络（conv 3×3: 尺寸为3×3大小的卷积核；relu: 激活函数；BN: 批归一化层；maxpool 2×2: 尺寸为2×2大小的最大池化层；maxunpool 2×2: 尺寸为2×2大小的最大反池化层; skip connection: 跳跃连接，将编码信息与解码信息相加；sigmoid: 激活函数，将输入值映射到0～1的概率）
Figure 2. Unet ghost imaging edge detection network (conv 3×3: convolutional kernel of size 3×3; relu is the activation function; BN: batch normalization layer; maxpool 2×2: maximum pooling layer of size 2×2; maxunpool 2×2: maximum inverse pooling layer of size 2×2; skip connection: jump connection to sum the encoded information with decoded information; sigmoid: activation function, maps input values to 0−1 probabilities)
图 4 Unet鬼成像边缘提取网络的训练损失变化过程和学习率变化曲线
Figure 4. Training loss variation process and learning rate variation curve of Unet ghost imaging edge extraction network
图 5 测试集的Unet鬼成像边缘提取网络输出边缘图和非极大值抑制算法处理后的边缘图
Figure 5. Output edge images of Unet ghost imaging edge detection network for the test set and the edge images processed by the non-maximal value suppression algorithm
图 6 多个字母或数字对应的Unet鬼成像边缘提取网络输出图
Figure 6. Output images of Unet ghost imaging edge detection network corresponding to multiple letters or numbers
图 7 字母K和数字7通过Unet鬼成像边缘提取网络输出的边缘结果与通过散斑移位鬼成像方法输出的边缘结果的对比
Figure 7. Comparison of the edge results output by the Unet ghost imaging edge detection network for the letter K and the number 7 with those output by the scatter-shift ghost imaging method
图 8 采样次数不同时，字母K和数字7通过Unet鬼成像边缘提取网络输出的边缘结果与通过散斑移位鬼成像方法输出的边缘结果的SNR和SSIM指标对比
Figure 8. SNR and SSIM metrics of the edge results output by the Unet ghost imaging edge detection network compared to those output by the scatter-shift ghost imaging method as the number of samples increases for the letter K and the number 7
-
[1] Pittman T B, Shih Y H, Strekalov D V, et al. Optical imaging by means of two-photon quantum entanglement[J]. Physical Review A, 1995, 52(5): R3429. doi: 10.1103/PhysRevA.52.R3429 [2] Bennink R S, Bentley S J, Boyd R W. “Two-photon” coincidence imaging with a classical source[J]. Physical Review Letters, 2002, 89(11): 113601. doi: 10.1103/PhysRevLett.89.113601 [3] Shapiro J H. Computational ghost imaging[J]. Physical Review A, 2008, 78(6): 061802(R). [4] Ferri F, Magatti D, Lugiato L A, et al. Differential ghost imaging[J]. Physical Review Letters, 2010, 104(25): 253603. doi: 10.1103/PhysRevLett.104.253603 [5] Sun Baoqing, Welsh S S, Edgar M P, et al. Normalized ghost imaging[J]. Optics Express, 2012, 20(15): 16892-16901. doi: 10.1364/OE.20.016892 [6] 牟晓霜, 黎淼, 王玺, 等. 基于分块平滑投影二次重构算法的单像素成像系统[J]. 强激光与粒子束, 2022, 34:119002 doi: 10.11884/HPLPB202234.220190Mou Xiaoshuang, Li Miao, Wang Xi, et al. Single-pixel imaging system based on block smoothed projected quadratic reconstruction algorithm[J]. High Power Laser and Particle Beams, 2022, 34: 119002 doi: 10.11884/HPLPB202234.220190 [7] Yang Mochou, Wu Yi, Feng Guoying. Underwater environment laser ghost imaging based on Walsh speckle patterns[J]. Frontiers in Physics, 2023, 11: 1106320. doi: 10.3389/fphy.2023.1106320 [8] 杨莫愁, 吴仪, 冯国英. 水下鬼成像的研究进展[J]. 光学学报, 2022, 42:1701003 doi: 10.3788/AOS202242.1701003Yang Mochou, Wu Yi, Feng Guoying. Research progress on underwater ghost imaging[J]. Acta Optica Sinica, 2022, 42: 1701003 doi: 10.3788/AOS202242.1701003 [9] Zhao Chengqiang, Gong Wenlin, Chen Mingliang, et al. Ghost imaging lidar via sparsity constraints[J]. Applied Physics Letters, 2012, 101: 141123. doi: 10.1063/1.4757874 [10] Gao Xinyu, Mou Jun, Xiong Li, et al. A fast and efficient multiple images encryption based on single-channel encryption and chaotic system[J]. Nonlinear Dynamics, 2022, 108(1): 613-636. doi: 10.1007/s11071-021-07192-7 [11] Liu Xuefeng, Yao Xuri, Lan Ruoming, et al. Edge detection based on gradient ghost imaging[J]. Optics Express, 2015, 23(26): 33802-33811. doi: 10.1364/OE.23.033802 [12] Mao Tianyi, Chen Qian, He Weiji, et al. Speckle-shifting ghost imaging[J]. IEEE Photonics Journal, 2016, 8: 6900810. [13] Ren Hongdou, Zhao Shengmei, Gruska J. Edge detection based on single-pixel imaging[J]. Optics Express, 2018, 26(5): 5501-5511. doi: 10.1364/OE.26.005501 [14] Wang Le, Zou Li, Zhao Shengmei. Edge detection based on subpixel-speckle-shifting ghost imaging[J]. Optics Communications, 2018, 407: 181-185. doi: 10.1016/j.optcom.2017.09.002 [15] Ren Hongdou, Wang Le, Zhao Shengmei. Efficient edge detection based on ghost imaging[J]. OSA Continuum, 2019, 2(1): 64-73. doi: 10.1364/OSAC.2.000064 [16] 樊玉琦, 温鹏飞, 许雄, 等. 基于卷积神经网络的雷达目标航迹识别研究[J]. 强激光与粒子束, 2019, 31:093203 doi: 10.11884/HPLPB201931.180388Fan Yuqi, Wen Pengfei, Xu Xiong, et al. Research on radar target track recognition based on convolutional neural network[J]. High Power Laser and Particle Beams, 2019, 31: 093203 doi: 10.11884/HPLPB201931.180388 [17] Wu Heng, Wang Ruizhou, Zhao Genping, et al. Sub-Nyquist computational ghost imaging with deep learning[J]. Optics Express, 2020, 28(3): 3846-3853. doi: 10.1364/OE.386976 [18] Zhu Ruiguo, Yu Hong, Tan Zhijie, et al. Ghost imaging based on Y-net: a dynamic coding and decoding approach[J]. Optics Express, 2020, 28(12): 17556-17569. doi: 10.1364/OE.395000 [19] Yang Xu, Yu Zhongyang, Xu Lu, et al. Underwater ghost imaging based on generative adversarial networks with high imaging quality[J]. Optics Express, 2021, 29(18): 28388-28405. doi: 10.1364/OE.435276 [20] 邵延华, 冯玉沛, 张晓强, 等. 基于深度学习的光学元件表面疵病识别[J]. 强激光与粒子束, 2022, 34:112002 doi: 10.11884/HPLPB202234.220023Shao Yanhua, Feng Yupei, Zhang Xiaoqiang, et al. Using deep learning for surface defects identification of optical components[J]. High Power Laser and Particle Beams, 2022, 34: 112002 doi: 10.11884/HPLPB202234.220023 [21] Wu Heng, Zhao Genping, Chen Meiyun, et al. Hybrid neural network-based adaptive computational ghost imaging[J]. Optics and Lasers in Engineering, 2021, 140: 106529. doi: 10.1016/j.optlaseng.2020.106529 [22] He Xing, Zhao Shengmei, Wang Le. Handwritten digit recognition based on ghost imaging with deep learning[J]. Chinese Physics B, 2021, 30: 054201. doi: 10.1088/1674-1056/abd2a5 [23] Bromberg Y, Katz O, Silberberg Y. Ghost imaging with a single detector[J]. Physical Review A, 2009, 79: 053840. doi: 10.1103/PhysRevA.79.053840 [24] Creswell A, Arulkumaran K, Bharath A A. On denoising autoencoders trained to minimise binary cross-entropy[DB/OL]. arXiv preprint arXiv: 1708.08487, 2017. [25] Wang Zhou, Bovik A C, Sheikh H R, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE Transactions on Image Processing, 2004, 13(4): 600-612. doi: 10.1109/TIP.2003.819861 -
下载:
点击查看大图
计量
- 文章访问数: 294
- HTML全文浏览量: 113
- PDF下载量: 43
- 被引次数: 0
