Multiple side-coupled images recognition in plastic optical fibers based on deep learning

Abstract

In recent years, the technology of multi-mode fiber (MMF) output speckle recognition has been widely used in the fields of optical imaging, endoscope devices, and fiber optic sensor. However, previous studies focused on the single-input transmission mode on a single fiber, which limited the application range of MMF. Meanwhile, the fiber optic sensors based on speckle recognition often lack multiplexing ability. In this paper, multiple lateral sections of plastic optical fiber (POF) are polished, and the images corresponding to 26 letters are coupled into POF form them. When images couple into POF from one lateral polishing section, by analyzing the output speckles based on deep learning technology, the accuracy of speckles recognition with convolution neural network (CNN) reaches over 95% and the structural similarity (SSIM) of reconstructed speckles with U-Net reaches up to 0.96. When two images couple into POF from different lateral polishing sections simultaneously, the output speckle can be reconstructed to the two images respectively, and the average SSIM value for them is 0.95. The experiment proves that the images coupled into fiber through the lateral polishing section of POF can be recognized and reconstructed well by output speckle analysis. This technology can promote the multiplexing ability of the speckles-recognition based fiber sensors or be applied to the field of fiber endoscopes.