Machine learning methods use nonlinear optics and structured light to expand the accuracy and capacity of information networks

By combining spatial dimensions and multiple degrees of freedom, structured light can greatly improve information capacity. In recent years, the combination of structured light patterns with image processing and machine intelligence has shown great potential for development in areas such as communications and sensing.

One of the most notable features of structured light fields is their 2D and 3D distribution of amplitude information, which can be effectively integrated with mature image processing techniques and even realize cross-media information transfer thanks to the currently transformative machine learning techniques.

Complex structured light fields based on coherent superposition states can carry abundant spatial amplitude information, and by further combining spatial nonlinear transformations, a significant increase in information capacity can be realized.

Zhang Zilong of Beijing Institute of Technology and Shen Yijie of Nanyang Technological University, together with their team members, proposed a new method to enhance information capacity based on complex-mode coherent superposition states and their spatial nonlinear transformation. By integrating machine vision and deep learning techniques, they realized large-angle point-to-multipoint information transmission with low bit error rate.

The study has been published in the journal Lasers and Photonics Review.

In this model, a Gaussian beam is used to obtain spatially nonlinear transformation (SNC) of structured light through a spatial light modulator, and a convolutional neural network (CNN) is used to identify the intensity distribution of the beam.

Comparing the fundamental superposition mode and the SNC mode, it is observed that as the order of the constituent eigenmodes of the fundamental mode increases, the encoding ability of the HG superposition mode is significantly improved over that of the LG mode, and the mode encoding ability after the spatially structured nonlinear transformation is significantly improved.

To verify the encoding and decoding performance based on the above model, a 50×50 pixel color image was sent. The RGB dimension of the image was divided into five chromaticity levels, consisting of a total of 125 kinds of chromaticity information, each encoded in 125 HG coherent superposition states. Furthermore, various degrees of phase jitter induced by atmospheric turbulence were loaded into these 125 modes via a DMD spatial light modulator and trained with deep learning techniques to form a dataset.

In addition, the analysis of the decoding effect of larger capacity was implemented using nonlinear transformation, and 530 SNC modes were selected for experimental measurement of the confusion matrix to these modes by convolutional neural network. Experimental results show that with clearer structural features, the SNC mode can significantly increase the data capacity while ensuring a similarly low bit error rate, and the data recognition accuracy can reach up to 99.5%.

In addition, the experiment also verified machine vision pattern recognition capabilities under diffuse reflection conditions, achieving simultaneous high-precision decoding using multiple receiving cameras with an observation angle of up to 70°.