A Comparative Study of Pretrained Convolutional Neural Network Model to Identify Plant Diseases on Android Mobile Device

Abstract

Optimization of mobile devices to be used as early identification tools of plant diseases using an application based on Convolutional Neural Network (CNN), with a high degree of accuracy and low power consumption is the focus of this study. The study was conducted using a dataset consisting of 38 different classes of PlantVillage dataset, which were then expanded using 2 classes of coffee plants and 4 classes of rice plants. The models that are going to be tested and compared consists of MobileNet V2, NasNet Mobile, DenseNet 121 layer, and InceptionV3. In the experiments, it was found that there was a decrease in accuracy when the application was run on a mobile device when compared to when it was run on a PC. Experiments also show that InceptionV3 is the most stable model and reaches the highest level of accuracy, which is 98.45% on mobile devices. However, InceptionV3 consumes a lot of resources when used on mobile devices. Meanwhile, MobileNet V2, NasNet Mobile, dan DenseNet 121, do not consume a lot of resources when tested on mobile devices. In terms of accuracy, NasNet Mobile reached 97.31%, then MobileNet V2 reached 96.55%, and DenseNet 121 reached 96.21%. Based on the research criteria, it can be concluded that the CNN model that is most suitable to be used on mobile devices is NasNet Mobile. Because it has a high degree of accuracy with low resource consumption.