Feature Extraction Enabled Deep Learning From Specklegram for Optical Fiber Curvature Sensing

Abstract

A novel approach for multimode fiber (MMF) curvature sensing based on the texture feature of the specklegram from the fiber facet is proposed, developed and demonstrated. Different from previously reported methods, the texture feature vector of the fiber specklegram is used as the descriptor of the curvature of the fiber. After the specklegrams from two kinds of MMF under different curvatures are recorded, the texture feature vector of each specklegram is first extracted using the uniform local binary pattern algorithm. Then, the texture feature vector is mapped into the target curvature in a nonlinear manner using a one-dimensional convolutional neural network (1D CNN). In the experiment, the model trained by the texture features of fiber specklegrams with a core diameter of 50 μ m exhibits high prediction accuracy and good generalization ability in the curvature range of 1.55-6.93 m-1. In the first experiment, good recognition results are provided by the texture feature of the MMF specklegram, for which the average curvature recognition accuracy is 100%. In the second experiment, the prediction error for 91.4% of the samples in the test set is within a range of ± 0.1,m -1. In addition, this approach has significantly better prediction ability and robustness than the traditional approach, which processes the specklegram image directly using a two-dimensional convolutional neural network (2D CNN) for sensing. The experimental results demonstrate the fiber curvature sensing capability of the proposed method based on the analysis of texture feature using a 1D CNN.