Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 4, 490(2024)
Design and implementation of multi-task diffraction neural network system
Fig. 1. Schematic diagram of digital full connection model and diffraction connection model
Fig. 2. Schematic diagram of the working principle of diffraction neural network
Fig. 3. Classification training results of four-layer（512×512）diffraction neural network image
Fig. 4. Phase parameters，angular spectrum and optical field distribution of each diffraction layer in network.
Fig. 5. Computational process of 4-layer diffraction neural network image recognition
Fig. 6. Simulation recognition results of 10 types of digital images
Fig. 7. Diffractive optical neural network model with nonlinear activation
Fig. 8. Phase distribution of diffraction layer of nonlinear diffraction neural network
Fig. 9. Performance comparison of 4-layer diffraction neural network
Fig. 10. Training results of diffraction neural network Fashion-MNIST
Fig. 11. Demonstration of nonlinear activation effect of phase modulation layer output and output layer of diffraction neural network
Fig. 14. Comparison of ZEMAX simulation input and CMOS optical plane simulation calculation output
Fig. 15. Diffraction neural network optical path system physical and diffraction network working area selection and design
Fig. 16. Image classification performance of 4-layer diffractive optical neural network
Table 1. LC-SLM parameter specifications
View tableView in ArticleTable 1. LC-SLM parameter specifications
参数规格 参数值 相位型SLM 振幅型SLM 生成厂商 上海瑞立柯（UPOlabs） 型号 HDSLM80R 工作波长 420～1 100 nm 分辨率 1 920×1 200 像素间距 8 μm 填充因子 >95% 调制位数 10 bit（1 024） 8 bit（256） 尺寸 110 mm×72 mm×24 mm 158.9 mm×110.5 mm×31 mm 相位范围 5.7π/532 nm 线性灰度级
Table 2. Performance of diffraction neural network with different number of diffraction neurons
View tableView in ArticleTable 2. Performance of diffraction neural network with different number of diffraction neurons
相位模板尺寸 衍射元尺寸/μm 网络层数 Epoch 训练正确率/% 测试正确率/% 64×64 8 4 100 19.2±0.5 19.0±0.1 128×128 8 4 100 58.5±0.5 49.8±0.5 256×256 8 4 100 75.6±0.5 75.5±0.5 512×512 8 4 100 87.6±0.5 87.5±0.5
Table 3. CMOS parameter specifications
View tableView in ArticleTable 3. CMOS parameter specifications
参数规格 参数值 生成厂商 大恒图像 型号 MER2-2000 分辨率 5 496×3 672 帧率 19.6 像元尺寸/μm 2.4 像元深度/bit 8，12 光谱 黑白 信噪比/dB 42.08
Get Citation
Copy Citation Text
Zirong WANG, Xingxiang ZHANG, Yongji LONG, Tianjiao FU, Mo ZHANG. Design and implementation of multi-task diffraction neural network system[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(4): 490
Category: Research Articles
Received: Apr. 15, 2023
Accepted: --
Published Online: May. 28, 2024
The Author Email: Xingxiang ZHANG (jan_zxx@163.com)
