Evaluation of CNN Models in Identifying Plant Diseases on a Mobile Device

Abstract

Farmers in rural areas with limitation of internet connectivity can be made possible for early plant diseases detection by using optimization of mobile devices which implemented an application based on Convolutional Neural Network (CNN) because of the computational efficiency. The researchers used a dataset containing 79 different classes of plant which was merged from several public domain datasets, which was evaluated and compared using four CNN models, consisting of MobileNetV3, EfficientNetB0, Mason model, and ShuffleNetV2. The experiment results showed that Mason model has a highest accuracy of 90.54% and the smallest output file of 0.85MB, MobileNetV3 88.83% with 16.85MB, EfficientNetB0 88.75% with 16.08MB, and ShuffleNetV2 83.52% with 15.89MB, which the four models have a slight accuracy decrease on both workstation and mobile devices. However, on resource consumption overall, MobileNetV3 consumed less than the others model, even though the value hasn't a huge difference of several points. It can be concluded that Mason model is the most suitable model to be implemented on mobile devices because of accuracy and low resource consumption.

A Hybrid ViT-CNN Model Premeditated for Rice Leaf Disease Identification

“…Subsequently, a fully connected CNN architecture is constructed in the second phase to facilitate automatic feature extraction and classification. Aristan and Kusuma [14] utilized a dataset containing 79 different plant classes, sourced from several public domain datasets, which they assessed and compared using four CNN models: MobileNetV3, EfficientNetB0, Mason model, and ShuffleNetV2. Results from the experiments indicated that the Mason model achieved the highest accuracy among the four.…”