Fig. 1. Schematic diagram of pseudo-thermal light correlation imaging^[37]

Fig. 2. Schematic diagram of CGI^[37]

Fig. 3. Traditional correlation reconstruction algorithms. (a) Reconstruction scheme based on IGI^[43]; (b) comparison of results between GI with singular value decomposition and other imaging methods^[46]

Fig. 4. Reconstruction algorithms based on compressed sensing. (a) Imaging results based on multi-scale illumination patterns^[47]; (b) comparison of imaging quality of multiple reconstruction algorithms based on randomly modulated matrix^[51]

Fig. 5. Reconstruction algorithms based on deep learning. (a) Simulation results of GI, CSGI and DLGI^[56], β represents sampling rate; (b) reconstruction results of end-to-end convolution neural network^[57]; (c) structural diagram of super-resolution convolution neural network based on compressed sensing^[60]; (d) GI reconstruction results calculated by deep learning framework based on dynamic decoding (Y-net)^[61]

Fig. 6. Optimization schemes of Hadamard matrix based on traditional correlation reconstruction algorithms. (a) Scheme of single-pixel imaging with spatial dynamic resolution^[63]; (b) Hadamard matrix sorting based on “Russian doll”^[64]; (c) scheme of multi-resolution progressive correlation imaging and imaging results ^[65]

Fig. 7. Optimization schemes of Hadamard matrix based on compressed sensing class reconstruction algorithms. (a) Diagram of illumination pattern based on the idea of origami^[69]; (b) Hadamard matrix sorting based on “cutting cake”^[70]; (c) four kinds of Hadamard matrix sorting^[71]; (d) comparisons of imaging quality for different kinds of illumination patterns at different sampling rates^[72]

Fig. 8. Energy sorting diagrams of Hadamard matrix^[73]

Fig. 9. Schemes of traditional reconstruction algorithms based on optimized orthogonal transformation matrix. (a) Illumination patterns of HSPI, 4-step FSPI and 3-step binary FSPI^[77]; (b) fast FSPI scheme via binary illumination^[78]; (c) computational weighted FSPI scheme via binary illumination^[79]

Fig. 10. Reconstruction algorithms based on deep learning. (a) Structure of deep convolutional auto-encoding network^[83]; (b) FSPI results based on deep neural networks^[84]; (c) FSPI results based on generating adversarial networks^[85]

Table 1. Performance comparisons of different reconstruction algorithms under random illumination patterns

Table 2. Comparisons of construction methods of different illumination pattern and reconstruction performance