Generative data augmentation for industrial object classification on edge tensor processing units
TL;DR
An experimental study shows that a cross-class accuracy of 0.842 in mineral construction and demolition waste classification was achieved with generative data augmentation, with significant improvement compared to a CNN trained only with original data, confirming the feasibility of edge TPUs and the effectiveness of generative data augmentation.
Abstract
Deep learning (DL) is widely used for object classification in industrial applications such as inspection, quality assurance, and safety monitoring. Although convolutional neural networks (CNNs) achieve high accuracy, their use in practice is limited by two major challenges: small, imbalanced datasets and the high energy consumption of GPU-based systems. These issues are especially critical in industrial environments. This work proposes a framework for industrial object classification under limited training data and energy constraints. In this framework, small datasets are expanded using generative data augmentation for the training of deep learning models, and the trained models are quantized to 8-bit and then implemented on edge tensor processing units (TPU) with low power consumption. An experimental study shows that a cross-class accuracy of 0.842 in mineral construction and demolition waste (CDW) classification was achieved with generative data augmentation, with significant improvement compared to a CNN trained only with original data. This confirms the feasibility of edge TPUs and the effectiveness of generative data augmentation.