TinyML Induced Portable Food Defect Detection for Edge Computing

Abstract

Food defect detection is a critical link in ensuring food safety. Traditional machine vision-based detection systems rely on cloud servers to complete algorithmic inference, suffering from problems such as large detection equipment volume and insufficient real-time performance. This paper proposes a portable lightweight detection scheme based on Terminal Machine Learning (TinyML) and Mobile Edge Computing (MEC). Through lightweight neural network model compression technology and edge node task collaboration mechanisms, low-power operation and millisecond-level response of the detection equipment are achieved. Experimental results show that in the scenario of fruit surface defect detection, the system achieves a detection accuracy of (Formula presented.), with a single-frame inference power consumption of only 85 mW, meeting the practical application requirements of portable devices.