An Analysis of Lightweight Convolutional Neural Network Models for Image Classification Task on Edge Device

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 analyze the recent studies regarding the lightweight CNN models for image classification tasks on edge devices. For that, the study objective is to investigate the current lightweight CNN models that are best for image classification task on edge device while the second objective is to evaluate and determine the most feasible model to be deployed on edge device (i.e., NVIDIA Jetson Nano). As methodology, several models were investigated and selected including a few variants from MobileNet, ShuffleNet, and VGG16 for performance analysis. The approach followed by transfer learning with training and validation were done on a computer and their inference were deployed and ran on Jetson Nano. In addition, the dataset used for the transfer learning is obtained from a dataset of image classification for different classes of leaf diseases for a groundnut plant. Then, performance evaluation according to training, validation, and inference results shows that the MobileNet variants were on the lower performance compared to VGG16 and ShuffleNetV2 variants. In addition, ShuffleNetV2 has better performance with 98.06% classification accuracy, 9.41 ms of inference time and 106.76 FPS. Thus, for future research, ShuffleNetV2 can be further optimize and modify so that it can be better support and generalization for various edge applications.