Comparative study on the performance of deep learning implementation in the edge computing: Case study on the plant leaf disease identification

Abstract

This research examines and explores four different pre-trained CNN deep learning models (AlexNet, VGG16, ResNet50, and DenseNet121) to be adopted as edge solution. The model is developed and evaluated using the PlantVillage dataset. Image transformation techniques and down sampling were carried out to mitigate the unbalanced class distribution problem. From the preliminary works, DenseNet121 was selected as the implementation model since it had the best accuracy (96.4%) in comparison with other models. The model is then tested on different endpoint devices (CPU, GPU and VPU) in different programming environments (standard PyTorch and OpenVINO) to test the consistency of models that work across different hardware and software configurations. Results show that the adoption of model on the edge (VPU) is able to maintain relatively high recall values, precision and F1 scores. The successful application of the model on VPU device shows the potential of the model to be implemented for the detection of plant diseases as edge solution/embedded application. • Plant disease identification deep learning model has been developed. • Model performance in different hardware and environment settings has been compared. • The model in edge setting with VPU has a better accuracy and inference time.