Fig. 1. Imaging process of computational optical imaging system

Fig. 2. Classification of typical deep learning based computational imaging techniques according to their objectives and motivations

Fig. 3. Single-frame lensless phase recovery using deep learning^[4]

Fig. 4. Fast FPM imaging with few images using deep learning technology^[10]

Fig. 5. Principle of fringe analysis method based on deep learning and comparison of phase reconstruction results^[37]. (a)Principle of fringe analysis method based on deep learning; (b) reconstruction result of FT; (c) reconstruction result of WFT; (d) reconstruction result of proposed deep-learning method; (e) reconstruction result of 12-step phase-shifting profilometry

Fig. 6. Framework of single-pixel technique using deep neural network^[14]

Fig. 7. Network of deep learning based imaging through scattering medium^[28]

Fig. 8. Basic framework of 3D diffraction tomography reconstruction based on deep learning^[26]

Fig. 9. Schematic of network of optical diffraction tomography based on deep learning^[27]

Fig. 10. Framework of defocusing distance calculation in digital holography based on deep neural network^[6]

Fig. 11. Schematic of automatic boundary segmentation framework for retinal OCT image^[23]

Fig. 12. Network framework of super-resolution imaging based on deep learning^[20]

Fig. 13. Experimental results of STED super-resolution imaging based on deep learning^[17]

Fig. 14. Results of imaging using very weak light based on deep learning^[34]. (a) Camera output with ISO 8000; (b) Camera output with ISO 409600; (c) recovered result from raw data of Fig. 14(a)

Fig. 15. Network framework of virtual staining imaging based on deep learning^[35]

Fig. 16. Model-driven deep-learning approach^[98]

Fig. 17. Causal hierarchy structure relevant to physics (left) and image classification (right)^[99]

Fig. 18. After adding slight noise into Panda image, CNN model recognizes image as Gibbon^[104]

Fig. 19. Comparison between deep learning and classical theoretical algorithm should be objective

Fig. 20. Forecasting earthquake using deep learning hit with rebuttals has been questioned