Enhancing Digit Recognition for Luminous Images in Edge Computing Through Transfer Learning With Robustness and Fault Tolerance

• Published in: IEEE transactions on reliability pp. 1 - 9
• Main Authors: Hsu, Tse-Chuan, Tsai, Yao-Hong, Cheng-Chung Chu, William
• Format: Journal Article
• Language: English
• Published: IEEE 06.06.2024
• Subjects: Computational modeling; Data models; Deep learning; digit recognition; Edge computing; Image recognition; incremental learning; Keras; Text recognition; Training; Transfer learning
• ISSN: 0018-9529; 1558-1721
• DOI: 10.1109/TR.2024.3393424

Deep learning is developing rapidly, and the emergence of many network architectures has brought significant breakthroughs to training recognition models. Due to the maturity of edge computing technology, we can perform regional image training through distributed nodes, which significantly improves the training model's accuracy while performing transfer learning to achieve better performance. In image processing technology, high-precision recognition of non-luminous images can currently be achieved by modeling, if we replace the visual recognition target with a glowing digital panel, the recognition rate cannot be the same as the static text recognition rate. This article uses Keras to build a convolutional neural networks deep learning model to identify glowing light-emitting diodes (LED) digits, incremental learning to complete transfer learning on edge computing nodes, and an integrated IoT architecture to achieve better recognition results. In the experiment, the verification results obtained from the distributed training nodes were successfully combined to model and retrain the nodes. The proposed distributed learning method can increase the accuracy from 70% to 89%. At the same time, the misclassified images can be retrained by integrating the transfer learning model with the distributed learning results, and the accuracy reaches more than 92%.