An Edge Intelligence-based Generative Data Augmentation System for IoT Image Recognition Tasks

Abstract

To solve the problem of data scarcity in IoT image recognition tasks, an EI-based generative data augmentation system is designed in this paper. The system adopts hybrid architecture, and edge server and cloud data center participate in computing together, which is logically divided into the training phase and running phase. The training phase completes data augmentation of source data and training of Convolutional Neural Networks (CNNs), while the running phase processes information through the pretrained CNNs, and completes iteration of the CNNs through expert review and self-learning mechanism. It is worth mentioning that a generative data augmentation model, an Effective Deep  Convolutional Generative Adversarial Network (E-DCGAN), has been proposed in the system. The experiments show that E-DCGAN is superior to the baseline model in image generation and data augmentation in both agricultural and medical fields. Compared with the baseline model, the FID values were reduced by 4.73% and 19.59%. Meanwhile, the use of E-DCGAN for data augmentation can significantly improve the image classification model (VGG19, AlexNet, ResNet50), and the average accuracy of agricultural and medical classification results has increased by 0.96% and 1.27% over the baseline.