Structured light can signi­ficantly enhance information capacity, due to its coupling of spatial dimensions and multiple degrees of freedom. In recent years, the combi­nation of structured light patterns with image processing and machine intelli­gence has shown vigorous develop­ment potential in fields such as communi­cation and detection. One of the most notable features of structured light field is the two and three-dimensional distri­bution of its amplitude information. This feature can effectively integrate with maturely developed image processing techno­logy and can also achieve cross-medium information transmission by virtue of machine learning technology currently driving profound changes.

Complex structured light field based on coherent superposition states can carry abundant spatial amplitude information. By further combining spatial nonlinear conversion, signi­ficant increases in information capacity can be realized. Zilong Zhang from Beijing Institute of Technology and Yijie Shen from Nanyang Techno­logical University, along with their teams’ members, proposed a new method for enhancing information capacity based on complex mode coherent super­position states and their spatial nonlinear conversion. By integrating machine vision and deep learning techno­logies, they achieved large-angle point-to-multipoint information trans­mission with low bit error rate. 

In this model, Gaussian beams are used to obtain spatial nonlinear conversion (SNC) of structured light through a spatial light modulator. Convolu­tional neural networks (CNN) are used to identify the intensity distribution of the beams. By comparing the basic super­position mode and the SNC mode, it is observed that with the increase in order of constituent eigen­modes of the basic mode, the encoding capability of HG super­position mode is significantly better than LG mode, and mode encoding capacity after spatial structured nonlinear conversion can be signi­ficantly improved.

To verify the encoding and decoding performances based on the above model, a 50×50-pixel color image was transmitted. The RGB dimensions of the image were divided into 5 chromati­city levels, comprising a total 125 kinds of chromaticity information, each encoded by 125 HG coherent super­position states. Addi­tionally, different degrees of phase jitter caused by atmospheric turbulence were loaded onto these 125 modes through a DMD spatial light modulator and trained with deep learning techno­logy to form a dataset. Further using nonlinear conversion, the analysis of higher capacity decoding effects was imple­mented, in which 530 SNC modes were selected for experimental measurement of the confusion matrix to these modes by convo­lutional neural networks.

The experimental findings indicate that due to more distinct structural features, SNC modes can still ensure similarly low bit error rates while signi­ficantly increasing data capacity, with a data recognition accuracy up to 99.5 %. Additionally, the experiment also verified the machine vision pattern recognition capability under conditions of diffuse reflection, achieving simul­taneous high-precision decoding by multiple receiving cameras with observation angles up to 70°. (Source: LPC-CAS)

Reference: Z. Zhang et al.: Spatial Nonlinear Conversion of Structured Light for Machine Learning Based Ultra-Accurate Information Networks, Laser Phot. Rev. 18, 2470039 (2024); DOI:  10.1002/lpor.202470039

Link: School of Optics and Photonics, Beijing Institute of Technology, Beijing, China