Fig. 1. Schematic of GRENOUILLE setup. (a) Top view; (b) side view

Fig. 2. Measurement experiment setup for SHG-FROG trace diagram

Fig. 3. Partial results measured by GRENOUILLE (from left to right are SHG-FROG trace diagrams, time domain inversion results, and frequency domain inversion results). (a) Central wavelength is 1030 nm, pulse width is 220 fs; (b) central wavelength is 1030 nm, pulse width is 360 fs; (c) central wavelength is 700 nm, pulse width is 350 fs; (d) central wavelength is 1030 nm, pulse width is 585 fs

Fig. 4. Complete process of proposed algorithm

Fig. 5. Effective and ineffective SHG-FROG trace diagrams (from left to right are SHG-FROG trace diagrams, retrieved trace diagrams, and algorithm inversion results). (a) Effective trace diagram; (b) ineffective trace diagram, exceeding the instrument's range; (c) ineffective trace diagram caused by excessive noise

Fig. 6. Pulse reconstruction structure using an optimized MO-ResNet

Fig. 7. Reconstruct results of SHG-FROG trace diagram. (a) SHG-FROG trace diagram; (b) temporal intensity and phase reconstructed by RANA; (c) temporal intensity and phase inverted by MO-ResNet

Fig. 8. SHG-FROG trace diagrams adding salt and pepper noise with different intensities

Fig. 9. Training loss of different models with training rounds. (a) Trace diagram recognition model; (b) MO-ResNet reconstruction model

Fig. 10. Performance comparison of pulse reconstruction algorithm on different trace diagrams. Original trace diagrams with actual pulse width of (a1) 284 fs, (a2) 300 fs, (a3) 408 fs, (a4) 515 fs; reconstruction results with actual pulse width of ( b1) 284 fs, (b2) 300 fs, (b3) 408 fs, (b4) 515 fs

Fig. 11. Comparison of reconstruction errors between two algorithms under different pulse widths