Transport-based pattern recognition versus deep neural networks in underwater OAM communications.

Area of Science:

• Optical Communications
• Machine Learning
• Image Processing

Background:

• Underwater optical communication systems face challenges due to signal attenuation.
• Orbital Angular Momentum (OAM) multiplexing offers potential for increased data capacity.
• Classifying OAM states in noisy, attenuated underwater environments is crucial for reliable communication.

Purpose of the Study:

• To compare machine learning and optimal transport (OT) based image classification methods for underwater OAM communications.
• To investigate the effectiveness of OT in attenuated water environments.
• To evaluate the performance of different classification algorithms on transformed OAM patterns.

Main Methods:

• Developed a model justifying optimal transport for attenuated underwater environments.
• Performed OAM pattern demultiplexing using optimal transport and deep neural networks.
• Applied the Radon cumulative distribution transform (R-CDT) to OAM patterns.
• Compared classification performance using nearest subspace algorithm, shallow CNN, and deep neural network on original and R-CDT transformed images.

Main Results:

• The Radon cumulative distribution transform (R-CDT) significantly improved OAM pattern classification accuracy compared to original images.
• The nearest subspace algorithm demonstrated superior performance over selected Convolutional Neural Networks (CNNs) for OAM pattern classification in underwater settings.
• Signal attenuation effects on OAM pattern classification were highlighted.

Conclusions:

• Optimal transport is a viable approach for OAM image classification in challenging underwater conditions.
• The R-CDT is an effective preprocessing technique for enhancing OAM pattern classification.
• Simpler algorithms like the nearest subspace method can outperform complex deep learning models in specific underwater communication scenarios.