Multimode Fiber Image Transmission via Cross-Modal Knowledge distillation

This is the authors' abstract. We don't add key points for this paper.

Image transmission through multimode fiber (MMF) has great potential to innovate the systems of endoscopic or neurological imaging. However, when injecting a light pulse with image information into an MMF, the spatial and temporal information is scrambled inside the MMF, leading to chaotic-like outputs of speckles and split pulses. Recently, the development of machine learning has enabled ultrafast MMF image transmission by reconstructing original images from the chaotic output pulses. Despite this, high-cost devices are required for pulse detection at the picosecond-level temporal resolution. Here, we propose a cross-modal knowledge distillation framework to address the device requirement on the customer side. Specifically, by leveraging a teacher model that is well-trained on high-quality data, the student model on the customer side can be trained with lower-quality data while maintaining a good performance.