Fig. 1. Process of detecting defocused speckle of irregular particles in IPI system

Fig. 2. Setup of simulated particle IPI system

Fig. 3. Autocorrelation analysis of irregular particle shape. (a) Simulated ice particle shape; (b) autocorrelation calculation result of (a); (c) defocused speckle in simulated particle IPI system; (d) 2D Fourier transform amplitude of (c) binarization result; (e) defocused speckle in experimental particle IPI system; (f) 2D Fourier transform amplitude of (e) binarization result

Fig. 4. Shape categories distribution and corresponding examples of ice particles in ICDC

Fig. 5. Training and testing processes of particle shape classification network

Fig. 6. Defocused speckle and correlation analysis of 5 ice particle shape categories

Fig. 7. Updating strategies of diffuser. (a) Rotating; (b) lateral translating; (c) longitudinal translating; (d) replacing

Fig. 8. Variation in speckle correlation after applying updating strategies

Fig. 9. Structure of DenseNet

Fig. 10. Training loss and accuracy curves of AlexNet, ResNet152, and DenseNet169

Fig. 11. Influence of defocused distance on classification accuracy. (a) Testing set classification accuracies at four different defocused distances; (b) confusion matrix of speckle testing set at defocused distance of 70 mm

Fig. 12. Speckle data at different bit depths. (a) 8 bit; (b) 2 bit; (c) 1 bit

Fig. 13. Influence of speckle bit depth on classification accuracy. (a) Testing set classification accuracies at different bit depths;(b) confusion matrix of 1 bit speckle testing set at defocused distance of 70 mm

Fig. 14. Influence of binarization threshold on classification accuracy of 1-bit speckle data. (a) Grayscale histogram of raw speckle and binarization thresholds. Inset: classification accuracies of speckle data under different thresholds; (b) 1-bit speckle and its sparsity corresponding to different grayscale thresholds

Fig. 15. Influence of speckle pattern size on classification accuracy of speckle data. (a) Speckle slices of different sizes; (b) testing set classification accuracies of different speckle pattern sizes

Table 1. Comparison of training parameters for AlexNet, ResNet152, and DenseNet169