Author Affiliations
1Key Laboratory of Light Field Manipulation and Information Acquisition, Ministry of Industry and Information Technology, Northwestern Polytechnical University, Xi'an, Shaanxi 710129, China2Shaanxi Key Laboratory of Optical Information Technology, School of Physical Science and Technology, Northwestern Polytechnical University, Xi'an, Shaanxi 710129, Chinashow less
Fig. 1. Simulation process. (a) Physical model; (b) forward problems fitting by neural network; (c) inverse problems fitting by neural network
Fig. 2. Simulation process. (a) Fully connected structure; (b) convolution operation
Fig. 3. Curve of activation function
Fig. 4. Main structure of network. (a) Backbone; (b) FCN; (c) U-net; (d) GAN
Fig. 5. Detail structure of network. (a) Residual block; (b) multi-scale block; (c) attention block, in which (c1) is channel attention and (c2) is spatial attention; (d) dense connected block
Fig. 6. Flow chart of network training and testing. (a) Training process; (b) testing process
Fig. 7. In-line holographic numerical reconstruction with CNN. (a) CNN is used to suppress “twin image” and autocorrelation artifacts
[32]; (b) end-to-end phase reconstruction using eHoloNet
[33] Fig. 8. Off-axis holographic numerical reconstruction with CNN. (a) U-net
[34]; (b) Y-Net
[36] Fig. 9. Applications of CNN in holographic reconstruction distance. (a) Regression model
[42-44]; (b) classification model
[47] Fig. 10. Fringe patterns analysis with CNN. (a) Method in Ref. [58]; (b) method in Ref. [59]; (c) method in Ref. [60]
Fig. 11. Phase unwrapping with CNN. (a) Method in Ref. [79]; (b) method in Ref. [80]; (c) method in Ref. [84]
Fig. 12. Ghost imaging technology. (a) Computational ghost imaging process; (b) ghost imaging reconstruction using neural network
Fig. 13. Computational ghost imaging with CNN. (a) DRU-Net
[99]; (b) DeepGhost
[101]; (c) DAttNet
[102] Fig. 14. Fourier ptychographic microscopy system
[116] Fig. 15. Applications of CNN in Fourier ptychographic microscopy. (a) Super-resolution reconstruction of complex amplitude lightfield
[115]; (b) aberration-free high resolution image reconstruction with pupil function estimation
[116]; (c) LED array position deviation correction to optimize reconstruction quality
[117] Fig. 16. Applications of CNN in super-resolution imaging. (a)Cross modal and super-resolution imaging with GAN
[124]; (b) end-to-end lensless microscopic super-resolution imaging
[127]; (c) hologram super-resolution optimization
[128] Fig. 17. Applications of CNN in scattering medium classification. (a) Network trained by synthetic data achieves classification in experimental application
[143]; (b) speckle pattern classification of face and non-face
[144] Fig. 18. Applications of CNN in scattering medium reconstruction. (a) Image reconstruction of speckle field behind optical fiber with CNN
[147]; (b) CNN is used to pre-reconstruct the phase
[148]; (c) CNN for image reconstruction in strong scattering media
[150] Fig. 19. Accurate choroidal segmentation using CNN
[157] Fig. 20. Optical fiber reconstruction in optical diffraction tomography using CNN. (a) Internal structure reconstruction of optical fiber with limited angle
[166]; (b) internal structure reconstruction of photonic crystal fiber with sparse angle
[167] Application field | Network structure | Loss function | Application problems |
---|
Digital holography | Backbone, U-net, GAN | MSE, MAE, cross entropy | Holographic reconstruction[30-39]: “twin-image” problem, “end to end” phase recovery, reconstruction of complex amplitude light fieldAuto focusing[42-49]: prediction of holographic reconstruction distanceOthers[50-54]: holographic image denoising, multi wavelength hologram fusion and reconstruction, reconstructed image enhancement | Fringe analysis | Backbone, FCN, U-net | MSE, MAE, regularization | Phase demodulation and 3D reconstruction[58-62,67]Fast recognition of fringes[63]Fringe image denoising[64-66] | Phase unwrapping | FCN, U-net, GAN | MSE, MAE, cross entropy, regularization | Phase unwrapping[79-86] | Application field | Network structure | Loss function | Application problems | Ghost imaging | Backbone, FCN, U-net, GAN | MSE, regularization | Noise suppression[94]Blind image reconstruction[95,97]Low sampling imaging[96,101-103,105]Lighting mode optimization[98,102-104] | Fourier ptychographic microscopy | U-net, GAN | MSE, MAE, regularization | Super resolution image reconstruction [112-113,115-116]Speed up reconstruction [113,115]Position deviation correction [117]Noise suppression[113-115] | Super resolution imaging | FCN, U-net, GAN | MSE, MAE, regularization | Super resolution imaging [120-131] | Scattering medium imaging | Backbone, U-net | MSE, MAE, cross entropy | Target classification [142-145]Image reconstruction [146-153]Modal decomposition of multimode fiber [154] | Optical tomography | Backbone, U-net, GAN | MSE, cross entropy, regularization | Coherence tomography [156-160]: high precision and fast image segmentation, image enhancementDiffraction tomography[164-167]: noise suppression, Inversion reconstruction |
|
Table 1. Applications of CNN in optical information processing