Performance comparison of MobileNet, EfficientNet, and Inception for predicting crop disease

Abstract

Disease plant food can cause significant loss in production agriculture since difficult to detect early symptoms of disease. Apart from that, the selection of Convolutional Neural Network (CNN) architecture for the detection of disease plants often faces the challenge of trade-offs between accuracy and efficiency. In this research, we propose a solution with compares the performance of three current CNN architectures, ie MobileNet, EfficientNet, and Inception, in context predictions of disease plant food. We implement a transfer learning approach to increase efficiency and performance model predictions. The contribution of this study is located on the guide practical for researchers and practitioners in choosing appropriate CNN architecture with need-specific application detection disease plant food. In this experiment, we use 3 datasets to represent plant food in Indonesia, namely rice, corn, and potatoes. Metric evaluation performances like accuracy, precision, recall, and F1-score are used to compare the results of the experiment. Experimental results show a significant difference in performance third tested architecture. MobileNet stands out in speed inference and necessity source low power, temporary EfficientNet shows a good balance between accuracy and efficiency. Inception delivers superior results in detecting feature complex however needs to source more power. In conclusion, the selection of CNN architecture for predictions of disease plant food must consider the trade-off between accuracy, speed inference, and necessity source power. These experimental results can give a guide valuable for practitioners in making appropriate technology with need Specific application detection disease plant food.