Chinese Optics Letters, Volume. 21, Issue 8, 080501(2023)
Suppressing defocus noise with U-net in optical scanning holography
Figures & Tables(19)
![OSH system setup[28]. BS1 and BS2, beam splitter; M1 and M2, mirror; AOFS, acousto-optic frequency shifter; p1 (x, y) and p2 (x, y), pupils; L1, L2, and L3, lens; PD, photo detector; BPF, band-pass filter.](/Images/icon/loading.gif){kind=icon}
Fig. 1. OSH system setup[28]. BS1 and BS2, beam splitter; M1 and M2, mirror; AOFS, acousto-optic frequency shifter; p1 (x, y) and p2 (x, y), pupils; L1, L2, and L3, lens; PD, photo detector; BPF, band-pass filter.
Fig. 2. The architecture of U-net. ‘Conv, 3 × 3’ represents a 3 × 3 convolution kernel with the ReLU activation function. ‘Padding=same’ means that the matrix dimensions of the input and output in the convolution layer are the same. ‘Maxpool 2 × 2’ represents the function to choose the maximum value from a 2 × 2 matrix. ‘Upsampling and conv, 2 × 2’ stands for upsampling using a 2 × 2 convolution kernel. Each blue box represents a multi-channel feature map, while the white ones represent the copied feature maps.
Fig. 5. The relationship between the loss function and iteration times.
Fig. 6. The reconstruction results with U-net. (a), (d), and (g) are the original images. (b), (e), and (h) are input images with speckle-like noise generated by different random phase pupils. (c), (f), and (i) are the corresponding reconstructed output images.
Fig. 7. Sectioning results with different noise ratios based on the U-net method.
Fig. 8. The original images for generating the training data sets.
Fig. 9. The relationship between the loss function and iteration times in the complex graphics.
Fig. 10. The test results of the complex graphics with U-net. (a),(d) are the original images of ‘Monkey’ and ‘Rice’. (b),(e) are the input images with speckle-like noise generated by different random phase pupils. (c),(f) are the corresponding output images of the U-net.
Fig. 11. (a) Object with two slices, and (b) the recorded hologram.
Fig. 12. (a) , (b) Sectional results of the conventional method. (c) , (d) Sectional results of the proposed U-net method, with z1 = 9 mm and z2 = 10 mm.
Fig. 14. (a)–(f) Sectional images of the 3D rocket. (g)–(l) Reconstructed images with the traditional method. (m)–(r) Reconstructed images with the proposed method.
Table 1. The Quantified Performance of the U-net Using Different Test Images
a View tableView in ArticleTable 1. The Quantified Performance of the U-net Using Different Test Images
a Test sample MSE PSNR (dB) SSIM (∈[0,1]) Input ‘ABC’ 1909.14 15.32 0.5782 VS ‘XYZ’ 1942.50 15.25 0.6171 Original ‘光’ 2138.30 14.80 0.5584 Output ‘ABC’ 15.8 36.15 0.9535 VS ‘XYZ’ 36.4 32.52 0.9326 Original ‘光’ 32.7 32.98 0.9383
Table 2. The Quantified Performance of the U-net with Different Noise Ratios
View tableView in ArticleTable 2. The Quantified Performance of the U-net with Different Noise Ratios
Test sample in Fig. 7 MSE PSNR (dB) SSIM (∈[0,1]) Input (a) 2256.34 13.35 0.3732 VS (b) 2387.33 12.33 0.3245 Original (c) 2489.95 10.58 0.2788 Output (d) 33.1 32.56 0.9305 VS (e) 32.1 32.47 0.9289 Original (f) 33.7 32.18 0.9245
Table 3. The Quantified Performance of the U-net Using Complex Test Images
View tableView in ArticleTable 3. The Quantified Performance of the U-net Using Complex Test Images
Test sample MSE PSNR (dB) SSIM (∈[0,1]) Input VS ‘Monkey’ 996.3 18.15 0.5613 Original ‘Rice’ 1218.2 18.15 0.5613 Output VS ‘Monkey’ 278.5 23.68 0.8127 Original ‘Rice’ 57.6 30.53 0.8378
Tools
Get Citation
Copy Citation Text
Haiyan Ou, Yong Wu, Kun Zhu, Edmund Y. Lam, Bing-Zhong Wang. Suppressing defocus noise with U-net in optical scanning holography[J]. Chinese Optics Letters, 2023, 21(8): 080501
Paper Information
Category: Diffraction, Gratings, and Holography
Received: Feb. 21, 2023
Accepted: Apr. 23, 2023
Posted: Apr. 24, 2023
Published Online: Aug. 9, 2023
The Author Email: Haiyan Ou (ouhaiyan@uestc.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20