Design and experimental research of orbitalangular momentum multiplexing holographybased on optical diffraction neural network

Thesis Information：

With the development of neural networks, the research of optical neural networks (ONN) has received wide attention. Starting from the basic theories of diffractive optics, scattered light, optical interference, and optical Fourier transform, researchers have successfully realized the optical linear operation of neural networks by using a variety of optical devices and materials, and further optimized the prediction and inference ability of ONN by introducing optical crystals, photovoltaic devices, and spatial light modulators to realize the optical nonlinear activation function, which has greatly facilitated the development of optical neural networks. Based on the flexible and programmable characteristics of the spatial light modulator, it provides a greater assistance for the optimization and experimental implementation of the optical path.

Fig.1. Schematic diagram of the experimental setup for OAM multiplexing holography based on ODNN. The spiral phase map and OAM multiplexing hologram are loaded into SLM1 and SLM2, respectively. After filtering and beam expansion, the coherent light is modulated into a vortex beam by SLM1, which is irradiated onto SLM2 to decode the corresponding target information in the OAM multiplexing hologram. Finally, the target image is reconstructed at the camera.

Fig.2. Schematic diagram of the physical process of OAM multiplexing holography based on ODNN. The optimized OAM multiplexing hologram is used to reconstruct the corresponding target image under different vortex light irradiation.

Fig.3. A neural network structure for OAM multiplexing holography based on ODNN.The physical process of OAM multiplexing holography is implemented as the forward propagation process of ODNN. During the training phase, the spiral phase and sampled target image are used as inputs and evaluation function labels for the neural network, respectively.Based on the output results of the network, the error backpropagation algorithm is used to iteratively optimize the neural network structure and the phase modulation values of neurons.Finally, an OAM multiplexing hologram that can reconstruct corresponding target images under different inputs is obtained at the hidden layer.

Fig.4. The overall design flowchart of OAM multiplexing holography based on ODNN, where the physical process in Fig. 1 is implemented as the forward propagation process of the neural network. In one iteration, a spiral phase map and a target image are input into the ODNN, and the OAM multiplexing hologram is updated through forward propagation, evaluation function calculation, and gradient descent. A training epoch consists of multiple iterations, and each spiral phase map and target image in the training dataset is input into ODNN to iteratively optimize the OAM multiplexing hologram. The complete training process consists of multiple epochs, and the convergence of the algorithm is determined by monitoring the average evaluation function value of each epoch, that is, whether the imaging quality no longer improves.

Fig.5. Simulation results of OAM multiplexing holography based on ODNN. (a) Normalized light intensity images of 36 targets reconstructed by holography under 36 different vortex beams, with the number in the lower right corner representing the corresponding topological charge l of the input vortex beam. (b) The normalized light intensity image of the digit "0" in holographic reconstruction, with the corresponding input spiral phase map in the lower right corner, and the inset at the bottom shows an enlarged image of a 29 × 29 pixel area around a single light spot in the reconstructed image, along with its intensity profile.

Fig.6. Comparison of imaging quality between the ODNN method and the classical method. (a-d) represent the comparison results of MSE, σim 2, η and SNR. Each indicator was simulated and calculated for 2-36 target quantities. The red and blue lines represent the indicator values of the classical method and the ODNN method, respectively. The yellow bar chart represents the percentage of performance improvement of the ODNN method compared to the classical method. The yellow dashed line represents the average performance improvement percentage across all target quantities.

Fig. 7. Experimental results of OAM multiplexing holography based on ODNN. The first row represents the spiral phase map of input l = 3, −3, 8, −8, 13, −13, the second row represents the normalized light intensity image of the corresponding simulated reconstructed target, and the third row represents the normalized light intensity image of the reconstructed target obtained in the experiment. The values in the upper left and lower right of the simulation and experimental results represent η (%) and SNR (in number), respectively

The parameters of the spatial light modulator used in this experiment are as follows:

In addition, we have launched the same high reflectivity version of the spatial light modulator FSLM-2K70-P02HR, which has a reflectivity of more than 95%.

2K×2K High Reflectivity & High Optical Utilization New Products

Recently, our company has re-launched the high reflectivity and high optical utilization series of reflective phase-type spatial light modulator FSLM-2K73-P02HR, with square large target surface design, high phase linearity, and high bit-depth, which enhances the optical utilization and improves the modulation precision at the same time, and continues to boost the development of scientific research and strive for excellence.

Written at the end:

Optical neural networks use optical systems to perform machine learning, and spatial light modulators, as an important optical field modulation device, have a natural advantage when applied to optical neural networks, offering great potential for parallel large-scale computation, ultra-low-power operation, and high-speed response. Optical neural networks, as a cutting-edge cross technology between the field of optics and the field of artificial intelligence, break through the technological limitations of traditional artificial neural networks, and are expected to be applied and developed in the fields of biomedicine, optical information communication, and machine vision.

 Article link: https://doi.org/10.1364/OE.538350