Perbandingan Metode Transfer Learning EfficientNet dan RegNet untuk Deteksi Penyakit pada Daun Tanaman Tomat
Muhammad Imron, Diyah Utami Kusumaning Putri, S.Kom., M. Sc., M. Cs.
2023 | Skripsi | ILMU KOMPUTER
Tomatoes are one of the most important crops in Indonesia. The total production of tomato plants in Indonesia is balanced with the total consumption of tomatoes by the Indonesian population. If there is a crop failure, meeting the needs of tomatoes must be balanced by importing them from other countries. Leaf disease is the main cause of tomato crop failure, therefore if leaf disease in tomato plants can be detected early then crop failure can be prevented. One of the alternatives in detecting disease on tomato plant leaves is CNN. Related research using CNN architectures such as AlexNet, VGG, and ResNet has been carried out a lot and has satisfactory results, but unfortunately this architecture uses a manual scaling process so it takes more experimentation time to get optimal performance. This can be overcome using the Neural Network Search algorithm which automates the scaling process. Several CNN architectures that use this algorithm, including EfficientNet and RegNet, have produced extraordinary performance.
This study aims to examine the performance of the EfficientNet and RegNet architectures in detecting diseases on tomato plant leaves. The object used in this study is a dataset of tomato leaf images from Kaggle. The dataset has 10 categories of 1100 images each. Before training, the image size will be resized and the image pixels will be normalized so that it fits the needs of the model. Then the dataset will be divided into training, validation, and test data with a ratio of 8:1:1. The augmentation process is also carried out using flip, rotation, shift and translation techniques to increase the number of images. The model used in this study is a pre-trained model that has been trained using data from ImageNet. The pre-trained model will then be fine-tuned to adjust the weight to the tomato leaf dataset. The training process will be carried out using different batch sizes and learning rates in order to obtain optimal performance. The model performance will then be evaluated using accuracy, precision, recall, and f1-score metrics.
This research generally includes four types of experiments that were trained using four different pre-trained models namely EfficientNetB0, EfficientNetV2S, RegNetX016, and RegNetY016. Overall, the four types of experiments produced satisfactory performance with figures above 97%. The model with the highest performance is RegNetX016 with a performance of 98.18% and the model with the lowest performance is EfficientNetB0 with a performance of 97.45%.
Kata Kunci : Penyakit Tomat, Artificial Intelligence, Machine Learning, Deep Learning, CNN, EfficientNet, RegNet