Robust deep learning-based detection and classification system for chipless Arabic RFID letters

Yousaf, J., Abed, A. M.A., Zia, H., Almajali, E., Tahir, F. A. and Rmili, H. (2023) Robust deep learning-based detection and classification system for chipless Arabic RFID letters. Engineering Applications of Artificial Intelligence, 122, 106147. (doi: 10.1016/j.engappai.2023.106147)

Abstract

This work demonstrates a novel approach for reliable and robust identification and detection of realized chipless RFID Arabic alphabets using deep learning (DL) method. The undertaken classification problem of Arabic RFID tags of various fonts and sizes requires a classification technique that can learn long-term dependencies. Hence, a Bi-Long Short-Term Memory (BiLSTM) model is developed to classify 28 chipless Arabic RFID letters of different font types and sizes using their back scattered dual-polarized radar cross section (RCS) characteristics. The RCS frequency response of each Arabic letter tag reflects its signature electromagnetic characteristics that vary with the change in its shape (variations in font type and size). Firstly, an RCS dataset of 28 Arabic alphabet tags with three different font types (Arial, Calibri, and Times New Roman) and 13 different font sizes (16 mm–28 mm with a step size of 1 mm) are generated using Finite-Difference Time-Domain (FDTD) method in the frequency range of 1–12 GHz (1001 steps). The dimensions of the resulting dataset are [28 (letters) × 13 (font sizes) × 1001 (frequency steps) × 2 (polarizations)] × 3 (font types). Multi-class classification of the frequency-series data of all realized 28 alphabet tags of various font types and sizes makes the problem challenging and novel. The developed BiLSTM model can accurately classify the particular letter tag with specific font type and size based on the optimized network with employed Leave-One-Out Cross-Validation (LOOCV). The achieved accuracy with only Arial ([(28 × 13 × 1001 × 2)]), Calibri ([(28 × 13 × 1001 × 2)]), Times New Roman ([(28 × 13 × 1001 × 2)]), and combined data set ([(28 × 13 × 1001 × 2)] × 3) is 75%, 74%, 75%, and 89% respectively. The proposed Bi-LSTM model is shown superior when compared to other methods such as SVM, decision trees, and KNN, as it classifies the data with much higher accuracy for the considered multi-class data. The obtained accuracies of the compared models are 6.4% (SVM), 17.30% (tree) and 27.4% (KNN) respectively, while the developed Bi-LSTM model with optimized hyperparameters achieved an accuracy of 96%.

• Item Type: Articles
• Keywords: Deep learning, alphabet RFID, chipless RFID, Arabic letter RFID, LSTM.
• Status: Published
• Refereed: Yes
• Glasgow Author(s) Enlighten ID: Tahir, Dr Farooq

Creator Roles

Tahir, F. A.: Visualization, Resources, Investigation, Writing – review and editing

• Authors: Yousaf, J., Abed, A. M.A., Zia, H., Almajali, E., Tahir, F. A., and Rmili, H.
• College/School: College of Science and Engineering > School of Engineering > Electronics and Nanoscale Engineering
• Journal Name: Engineering Applications of Artificial Intelligence
• Publisher: Elsevier Science
• ISSN: 0952-1976
• ISSN (Online): 1873-6769
• Published Online: 21 March 2023
• Copyright Holders: Copyright © 2023 Elsevier Ltd.
• First Published: First published in Engineering Applications of Artificial Intelligence 122:106147
• Publisher Policy: Reproduced in accordance with the publisher copyright policy