Search

In this paper, using a deep learning technique, the sensing characteristic of a refractive index (RI) sensor based on a multimode optical fiber speckle-gram has been extracted. The investigated sensor structure consists of a tapered optical fiber or microfiber (MF) induced on a standard multimode optical fiber (MMF). Due to the presence of evanescent waves around the MF, the guiding status of the end MMF is dependent on the external RI around the taper. Thereby, the propagating modes and resulting interference pattern or speckle-gram become sensitive to the changes in the taper environment. Herein, by analyzing and quantifying the speckle-gram modulation, the RI changes around the taper are characterized. To extract the sensing characteristic and improve its measurement accuracy, deep learning algorithms and specifically the ResNet deep neural network are used in the processing of the output speckle-grams. Using this algorithm, while the tapered fiber is immersed in different refractive index solutions in the range of 1.33 to 1.368, a set of output speckle images is captured for learning and then a number of them are used for measurement, and finally the accuracy of detecting the refractive index of the solutions RI has been achieved up to 95%. [ABSTRACT FROM AUTHOR]

According to statistics, over half a billion vehicles are moving on the roads worldwide. All vehicles have a license plate as their primary identifier, which is one of the most suitable tools for vehicle identification. Automatic License Plate Recognition (ALPR) can be effective in improving road security, reducing traffic congestion, enhancing transportation efficiency, preventing car theft, and more. Traditional methods proposed for license plate detection mainly relied on manual feature extraction and lacked could not be generalized to variable image components in different conditions. With recent advancements in deep learning, algorithms have emerged that can automatically extract high-level representations of images in addition to learning complex image structures. Therefore, in this paper, the high capacity of deep neural networks is utilized for learning license plate identifiers. The proposed model consists of two stages: license plate localization and plate identification. For localization, a combination of Convolutional Neural Network (CNN) and Generative Adversarial Network (GAN) is used in an encoder-decoder network. The proposed model is evaluated on two datasets, FZU Cars and Stanford Cars, and based on the experimental results, it outperforms baseline methods in terms of accuracy on both datasets. [ABSTRACT FROM AUTHOR]