12 December 2024 in Signal Processing
19 December 2024 in Optics Express
We propose a conformal vision transformer (CViT)-based demodulation for the perfect optical vortices shift keying (POV-SK) signal in the low-density parity check (LDPC) coded free-space optical (FSO) link. Despite the growing interest in POV for FSO links, atmospheric turbulence (AT) induces phase distortions, resulting in POV-SK demodulation errors and degrading POV-SK FSO communication performance. The CViT demodulator utilizes conformal mapping to reshape the circular POV-SK patterns into rectangles, enabling more efficient feature learning. Paired with ViT’s comprehensive global feature extraction capabilities, the CViT demodulator effectively counters turbulence and boosts accuracy in identifying the POV-SK modes. In this study, the CViT POV-SK demodulation method is thoroughly described. The extensive numerical simulations compare CViT, ViT, Conformal ResNet18, and the lightweight RCDT-based ’shallow’ CNN in terms of recognition accuracy, PSNR, and bit error rate (BER). Under severe turbulence (C
n
2=4.4×10−12m−2/3), conformal mapping can improve the recognition accuracy of ViT from 0.7636 to 0.8437 (CViT), while reducing the BER from 3.6 × 10−2 to 1.1 × 10−3. Under the premise of keeping recognition accuracy and BER the same, CViT significantly reduces the number of model parameters and time consumption compared to Conformal ResNet18, which is essential for meeting the system’s future real-time demands. In addition, in moderate turbulence, the performance of the RCDT-based CNN is acceptable, but it significantly lags behind CViT in severe turbulence, with the BER being an order of magnitude higher than that of CViT.
2 January 2026 in Optics Express
In free-space optical communication (FSOC) using wavefront-sensor-less (WFS-less) adaptive optics (AO), stochastic parallel gradient descent (SPGD) is commonly used to mitigate the adverse effects of atmospheric turbulence (AT) and correct the wavefront distortions. However, conventional SPGD often converges slowly and is prone to becoming trapped in local optima, especially under severe phase aberrations. In this study, we propose ViT-SPGD, a novel wavefront-correction method that integrates a vision transformer (ViT) with SPGD, to significantly improve wavefront correction performance. Specifically, the ViT predicts a deterministic, high-confidence gradient direction while SPGD provides stochastic gradient perturbations. By adaptively fusing the two directions, the hybrid update leverages ViT’s learned before accelerate convergence while preserving SPGD’s stochastic exploration whenever the ViT guidance is uncertain. We first detail the ViT-SPGD design and fusion mechanism, then validate its performance by comparing it with the original SPGD and two established variants, AdamSPGD and NSPGD. Simulation results, including the system-level dynamic simulations, show that ViT-SPGD substantially accelerates convergence and improves robustness in phase distortion correction. By combining deep learning techniques (i.e., ViT), with traditional iterative optimization (i.e., SPGD), ViT-SPGD provides a practical blueprint for intelligent, high-performance AO and helps pave the way for more robust FSOC operation in atmospheric conditions.
8 December 2024 in Optical Fiber Technology
2 September 2025 in Optics Letters
This Letter presents a new artificial intelligence (AI)-based technique for decoding 16/32-ary orbital angular momentum shift keying (OAM-SK). Departing from the conventional CNN-based method that involves directly receiving vortex light at the terminal and decoding it with a convolutional neural network, our approach employs a 5×5 Dammann vortex grating (DVG) to generate an OAM-SK light array (from −12th to +12th diffraction order) in the receiving plane. Using DVG, the OAM-SK beam’s light pattern evolves systematically across diffraction orders, transforming the optical array into a sequential signal for recognition by a convolutional neural network-long short-term memory (CNN-LSTM) model. To our knowledge, the DVG-CNN-LSTM method presents the first successful integration of deep learning model with DVG to improve OAM-SK decoding performance. The simulation results indicate that our innovative method markedly enhances recognition accuracy with the compact architecture with low computational overhead, an improvement attributed to the advanced acquisition of OAM-SK light pattern.