Abstract
This work demonstrates a computational method for predicting light propagation through a single multimode fiber (MMF) using a deep convolutional neural network (CNN). By leveraging a digital micro-mirror device (DMD) for spatial light modulation, we gathered aligned data pairs of input patterns and captured intensity-only speckle images. The sufficiently trained network demonstrates excellent performance in directly inferring fiber output, validated by mean squared error (MSE), correlation coefficient (corr), and structural similarity index (SSIM).
Introduction
Multimode fibers (MMFs) are essential in optical engineering for their high data capacity and potential in ultra-thin endoscopy. However, light propagation in MMFs is complicated by mode interference, dispersion, and coupling, which transform coherent input into chaotic speckle patterns. Traditional reconstruction methods, such as Transmission Matrix (TM) calculation or digital phase conjugation, are computationally intensive and require highly stable experimental setups. This paper presents an efficient, end-to-end deep learning approach to map the relationship between input patterns and output speckle intensity.
Experimental Setup and Methods
The experimental rig utilizes a laser beam modulated by a Digital Micro-mirror Device (DMD) to generate binary input patterns. This modulated light is coupled into an MMF, and the resulting speckle pattern at the distal end is recorded by a camera.
To process this data, we designed a deep convolutional neural network architecture. The model consists of three convolutional layers using $3 \times 3$ kernels to extract features, gradually increasing channel depth from 5 to 24. Regularization is achieved through Batch Normalization and Dropout layers to enhance training speed and prevent overfitting. A massive dense layer, containing nearly 200 million parameters, maps these features to the final $96 \times 96$ output pixels.
Datasets and Training
We generated a dataset of 7,780 aligned pairs of $36 \times 36$ binary input patterns and $96 \times 96$ intensity-only speckle images. After data cleaning to remove overexposed or dim images, 7,002 pairs were used for training and validation, with 778 pairs reserved for blind testing. The model was trained using the AdaDelta optimizer to minimize the Mean Squared Error (MSE) between predicted and actual speckle patterns.
Key Results and Discussion
The model achieved exceptional prediction performance across all validation standards. The average correlation (Corr) between predicted and actual speckle patterns reached 0.9770, while the Structural Similarity Index (SSIM) was 0.9315, indicating high fidelity in spatial structure preservation.
One of the most significant advantages of this approach is speed. Once trained, the network processes images in just 0.0224 seconds on a standard laptop CPU, enabling real-time inference at approximately 45 frames per second. This performance suggests that the network effectively learns an implicit physical mapping of the fiber's transformation without needing explicit Maxwell equations.
Conclusion
We have successfully demonstrated a deep learning framework capable of predicting light propagation through multimode fibers with high accuracy and speed. By working entirely with intensity-only images, this method simplifies hardware requirements compared to phase-retrieval techniques. Future work will focus on the model's robustness to dynamic fiber bending and scaling to higher-resolution imaging.
Summary updated