Author Affiliations
College of Information and Computer, Anhui Agriculture University, Hefei, Anhui 230031, Chinashow less
Fig. 1. Basic principle diagram of wavefront restoration based on CNN. (a) Diagram of optical path; (b) training and prediction data structure of CNN
Fig. 2. Diagram of CNN structure used for wavefront restoration
Fig. 3. Eight-layer CNN model for wavefront restoration
Fig. 4. Loss function curves in network training and testing processes. (a) D/r0=5; (b) D/r0=15; (c) D/r0=1-15
Fig. 5. Single-frame aberration randomly generated at D/r0=5 and its restoration results. (a) Wavefront to be restored; intensity images of (b1) focal plane and (b2) defocus plane; (c1) wavefront restored by CNNM1; (c2) residual aberration restored by CNNM1; (d1) wavefront restored by CNNM3; (d2) residual aberration restored by CNNM3
Fig. 6. Single-frame aberration randomly generated at D/r0=15 and its restoration results. (a) Wavefront to be restored; intensity images of (b1) focal plane and (b2) defocus plane; (c1) wavefront restored by CNNM2; (c2) residual aberration restored by CNNM2; (d1) wavefront restored by CNNM3; (d2) residual aberration restored by CNNM3
Fig. 7. Comparison between actual Zernike coefficient and predicted Zernike coefficients by CNNM1,CNNM2, and CNNM3. (a) D/r0=5; (b) D/r0=15
Fig. 8. Strehl ratio of axial light intensity before and after distortion wavefront compensation. (a) D/r0=5; (b) D/r0=15; (c) D/r0=1-15
Dataset No. | D/r0 | Training dataset | Test dataset |
---|
D/r0interval | Data volume/interval | Total datavolume | D/r0interval | Data volume/interval | Total datavolume |
---|
1 | 5 | — | 100 | 15000 | — | 10 | 1500 | 2 | 15 | — | 100 | 15000 | — | 10 | 1500 | 3 | 1-15 | 1 | 100 | 15000 | 1 | 10 | 1500 |
|
Table 1. Description of training and testing data volumes
Layer No. | Type | Image size | Filter size | Stride | Pading | Number of kernels |
---|
0 | Input | 224×224×2 | - | - | - | - | | Conv1 | 224×224×2 | 11×11 | 4 | - | 64 | 1 | Pooling1 | 55×55×64 | 3×3 | 2 | - | - | | ReLU | 55×55×64 | - | - | - | - | | Conv2 | 27×27×64 | 5×5 | 1 | 2 | 256 | 2 | Pooling2 | 27×27×192 | 3×3 | 2 | - | - | | ReLU | 13×13×192 | - | - | - | - | 3 | Conv3 | 13×13×192 | 3×3 | | 1 | 384 | | ReLU | 13×13×384 | - | - | - | - | 4 | Conv3 | 13×13×384 | 3×3 | | 1 | 384 | | ReLU | 13×13×384 | - | - | - | - | | Conv4 | 13×13×384 | 3×3 | | 1 | 256 | 5 | Pooling4 | 13×13×256 | 3×3 | 2 | - | - | | ReLU | 13×13×256 | - | - | - | - | 6 | Fully connected | 6×6×256 | - | - | - | - | 7 | Fully connected | 4096×1 | - | - | - | - | 8 | Tanh and output | N×1 | - | - | - | - |
|
Table 2. Parameters of CNN network model for wavefront restoration