Fig. 1. Schematic diagram. (a) Conventional imaging model; (b) computational imaging model

Fig. 2. Schematic of regression problem solved using neural network. (a) Training; (b) testing; (c) fitting process

Fig. 3. Diagram of data acquisition methods in computational imaging. (a) Images reconstructed by traditionally complex and costly methods; (b) presetting real objects with SLM; (c) data obtained by numerical simulation

Fig. 4. Schematic of fully connected neural network. (a) Single hidden layer; (b) multi hidden layers

Fig. 5. Simplified diagram of convolutional neural network. (a) Convolution: different convolution kernels corresponding to different feature maps; (b) pooling: replace data within the scope of operation with its maximum or mean; (c) deconvolution: interpolate data with zeros and then implement convolution; (d) convolution

Fig. 6. Main problems in neural network training. (a) Local minimum; (b) overfitting

Fig. 7. Flow chart of neural network training

Fig. 8. Applications of deep learning in scattering imaging. (a) Imaging through strong scattering media^[28,74]; (b) imaging through different ground glasses^[27]

Fig. 9. Applications of deep learning in digital holography. (a) Removing twin image^[32]; (b) end-to-end phase reconstruction^[34]; (c) end-to-end complex amplitude reconstruction^[66]

Fig. 10. Applications of deep learning in computational ghost imaging. (a) Improving signal-noise-ratio^[29]; (b) reconstructing objects by using intensity sequences^[31]

Table 1. Typical applications of deep learning in computational imaging