Image reconstruction through a hollow core fiber via deep learning

Highlights

• CNN is employed to reconstruct the images as objects pass through the HCF.

• Accuracy and PCC are used to evaluate CNN and fidelity of image reconstruction.

• The fidelity of the recovery images is good as high value of accuracy and PCC.

• The larger diameter of HCF is, the better is fidelity of reconstructed image.

• The longer length of HCF is, the better is fidelity of reconstructed image.

Abstract

A laser beam carrying object information will become a random speckle pattern as it propagates through a hollow core fiber (HCF) of large inner diameter. We propose and experimentally demonstrate a technique that employs the convolutional neural network (CNN), which is one type of deep learning, for image reconstruction from the random speckle pattern. In the experimental demonstration, 4000 speckle images and 4000 corresponding original images are used to train CNN, and 1000 speckle images are used to test the trained CNN. The experimental results demonstrate the possible realization of deep learning for the image reconstruction through the HCF. Furthermore, the influence of the diameter and length of HCF on the fidelity of the image reconstruction is quantitatively investigated by determining the accuracy and Pearson correlation coefficient (PCC).

Introduction

Optical fiber has a strong ability of information transmission, which makes it widely used in optical communication and imaging, etc.  [1], [2], [3], [4]. Majority of the optical fibers are fabricated with silicon inner core, which produces the material absorption, scattering, dispersion, and nonlinear effects in light propagation. Recently, the hollow core fiber (HCF) whose inner core is of air is utilized to eliminate the undesirable effects generated by the silicon core  [5], [6]. Therefore, HCFs are beneficial to the transmission of high-power mid-infrared laser  [7], and the fabricated HCF bundle is used for infrared thermal imaging  [8]. Besides, the HCFs are considered to apply in endoscopic imaging due to its high flexibility and nontoxicity  [6]. Moreover, HCFs filled with different liquids have several applications in sensing temperature, vibration, and humidity etc.  [9], [10]. However, similar to multimode fiber (MMF), light transmission through multimode HCF with large hollow inner core generates speckle pattern at the distal end of the HCF. In recent years, a number of techniques, such as digital phase conjugation and transmission matrix etc., have been developed in recovering images from the speckle pattern as objects pass through scattering media or MMF  [11], [12], [13]. More recently, deep learning technology has been successfully applied in the field of computational optical imaging  [14], three-dimensional imaging with optical tomography  [15], lensless computational imaging, etc.  [16]. Furthermore, convolutional neural network (CNN), as one subclass of deep learning, is efficiently demonstrated in image reconstruction in various optical fibers such as the MMF  [17], [18], [19], multicore fiber  [20], glass-air Anderson localized optical fiber  [21], and in imaging through scattering media  [22], [23]. The CNN can implicitly learn the forward operator and regularization function through the training process without knowing them in advance  [24], [25], [26], [27]. The trained CNN has an accurate computing architecture for realizing imaging from the speckle pattern generated by the transmission of light through scattering medium or MMF. The CNN utilizes a large set of matching input (original image) and output (speckle pattern) pairs to optimize the parameters and to build the optimal computing architecture. In this paper, we extend the computing potential of CNN to reconstruct high-resolution image through HCF. The image reconstruction scheme makes use of a training process employed with a large set of object–speckle pairs i.e., recorded speckle pattern corresponding to various objects transmitted through the HCF. In addition, the recovery of high-resolution images free from undesirable effects will open promising low-cost research applications in comparison with high-cost MMF.