Weight-location recognition in a plastic optical fiber using a convolutional neural network

Abstract

Convolutional neural networks (CNNs) have proven to be highly efficient in performing arduous tasks such as image reconstruction, image classification, etc. on multimode fiber (MMF) with intensity-based detection. In this work, a CNN model known as AlexNet has been employed for weight-location recognition analysis by performing an image classification task. A simple experiment is performed in which random weights in the range 0.5 kg-3 kg are applied on the plastic optical fiber (POF) at six unequally-spaced, pre-determined locations along the length of the POF for data acquisition comprising of specklegram images. The experiment is repeated for 1 m, 2 m and 3 m fiber length. These images are split into training, validation and test dataset in the ratio of 80 %, 10 % and 10 % respectively. The images in the training dataset are employed for training the model, whereas the validation dataset is used for validating the model. The model makes output predictions of weight-location on the test dataset with optimal recognition (classification) accuracy. The recognition accuracy of 100 % for 1 m and 2 m, whereas 99.8 % accuracy is achieved for 3 m. The results suggest that there is a low loss in recognition accuracy with the increase in fiber lengths. Hence, there is a negligible impact of fiber length on recognition accuracy.