Systematic review of a lightweight convolutional neural network architectures on edge devices

Abstract

A lightweight convolutional neural network (CNN) has become one of the major studies in machine learning field to optimize its potential for employing it on the resource-constrained devices. However, a benchmark for fair comparison is still missing and thus, this paper aims to identify the recent studies regarding the lightweight CNN architectures including the types of CNN, its applications, edge devices usage, evaluation types and matrices, and performance comparison. The preferred reporting items for systematic reviews and meta-analysis (PRISMA) framework was used as the main approach to collect and interpret the literature. In the process, 37 papers were identified as meeting the criteria for lightweight CNNs aimed at image classification or regression tasks. Of these, only 20 studies explored the use of these models on edge devices. To conclude, MobileNet appeared as the most used architecture, while the types of CNN focused on image classification for the general-purpose application. Following that, the NVIDIA Jetson Nano was the most utilized edge device in recent research. Additionally, performance evaluation commonly included measures like accuracy and time, along with metrics such as recall, precision, F1-Score, and other similar indicators. Finally, the average accuracy for performance comparison can serve as threshold value for future research in this scope of study.