Transfer learning convolutional neural network for weight classification in a plastic optical fiber

Abstract

Abstract In this work, a machine learning approach known as Convolutional Neural Network (CNN) has been proposed for classifying several weights applied on a plastic optical fiber (POF) based on specklegram (speckle pattern) images. We have optimized the CNN model using various inbuilt Keras optimizers namely Adam, Adamax, Nadam, and RMSprop in the python programming language. The best classification accuracy is shown by the RMSprop optimizer with 76.1%. Further, we have achieved improvement in accuracy by employing a transfer learning approach on pre-trained models namely VGG-16 and VGG-19. The classification accuracies of 84.2% and 84.5% are obtained for VGG-16 and VGG-19, respectively. It implies that there is an increase in classification accuracy of around ∼8% after implementing transfer learning. In order to assess the efficacy of the transfer-learned models, we have investigated the surrounding external vibrations applied to a section of the POF at three different frequencies, namely, 10 Hz, 5 kHz, and 50 kHz. We found that the optimal classification accuracy (>70%) is obtained till 5 kHz, beyond which the value decreases drastically. This methodology can lead to the development of real-time, smart sensors for weight detection.