Infrared and Laser Engineering, Volume. 51, Issue 8, 20220215(2022)
Deep learning-based image reconstruction through turbid medium （invited）
Figures & Tables(12)
Fig. 1. Schematic of experimental setup. L, lens; P, polarizer; BS, beam splitter; SLM, spatial light modulator; CCD, charge-coupled device
Fig. 2. pix2pix model diagram. (a) U-net architecture of the generator; (b) PatchGAN architecture of the discriminator. Conv, convolution; BN, Batch-Normalization
Fig. 3. Ground truth, speckle patterns and reconstructed images of different scattering media (Set 1-3). Values of SSIM are marked in the upper left corner of the reconstructed images
Fig. 4. Ground truth, speckle patterns and reconstructed images of polystyrene suspension with different concentrations (Set 3-6). Values of SSIM are marked in the upper left corner of the reconstructed images
Fig. 5. Comparison of SSIM, PCC and PSNR average values of Set 1- 6 restored images
Fig. 6. Ground truth, speckle patterns and reconstructed images of polystyrene suspension with different concentrations (Set 3-6). Values of SSIM are marked in the upper left corner of the reconstructed images
Fig. 7. Loss evolution of the training loss, ReLU_down: using activation function of ReLU, learning rate of the optimizer decays; LeakyReLU_down: using activation function of LeakyReLU, learning rate of the optimizer decays; ReLU: using activation function of ReLU, learning rate of the optimizer unchanged
Fig. 8. Hand drawn graffiti original image, speckle and its restored image collected from the calcium carbonate suspension. The values of SSIM are marked in the upper left corner of the reconstructed images
Table 1. Scattering medium added in the turbid water
View tableView in ArticleTable 1. Scattering medium added in the turbid water
Scattering medium Set 1 Sea salt Set 2 Calcium carbonate Set 3 Polystyrene (250 mg/10 mL, 2 mL) Set 4 Polystyrene (250 mg/10 mL, 1.5 mL) Set 5 Polystyrene (250 mg/10 mL, 1 mL) Set 6 Polystyrene (250 mg/10 mL, 0.5 mL)
Table 2. Values of SSIM, PCC, and PSNR of Set 1-6
View tableView in ArticleTable 2. Values of SSIM, PCC, and PSNR of Set 1-6
SSIM PCC PSNR/dB Set 1 0.951 0.940 18.733 Set 2 0.92 0.923 16.934 Set 3 0.886 0.829 14.814 Set 4 0.925 0.891 16.681 Set 5 0.939 0.933 18.614 Set 6 0.951 0.957 20.462
Table 3. Evaluation index of the reconstructed images of test set in different training methods
View tableView in ArticleTable 3. Evaluation index of the reconstructed images of test set in different training methods
SSIM PCC PSNR/dB ReLU_down 0.932 0.897 17.639 LeakyReLU_dwon 0.906 0.86 16.273 ReLU 0.928 0.893 17.564
Table 4. SSIM, PCC, and PSNR of Set 3-6 mixed used pix2pix and U-net, respectively
View tableView in ArticleTable 4. SSIM, PCC, and PSNR of Set 3-6 mixed used pix2pix and U-net, respectively
SSIM PCC PSNR/dB pix2pix 0.932 0.897 17.639 U-net 0.832 0.809 13.088
Tools
Get Citation
Copy Citation Text
Zhiyuan Wang, Xuetian Lai, Huichuan Lin, Fuchang Chen, Jun Zeng, Ziyang Chen, Jixiong Pu. Deep learning-based image reconstruction through turbid medium （invited）[J]. Infrared and Laser Engineering, 2022, 51(8): 20220215
Paper Information
Category: Special issue——Scattering imaging and non-line-of-sight imaging
Received: Mar. 22, 2022
Accepted: --
Published Online: Jan. 9, 2023
The Author Email: Chen Fuchang (chenfuchang@mnnu.edu.cn), Chen Ziyang (ziyang@hqu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20