The goal of this project was the application of different Convolutional Neural Network (CNN) architectures for the resolution of the plant disease and pest detection problem, based on the examples present in the PlantVillage dataset. For this, different ResNet and EfficientNet models were trained and compared based on various performance metrics, with these models being either trained fully from scratch on the previously mentioned dataset, or trained with transfer learning from other models pre-trained on either the broad ImageNet dataset, or a more specific plant disease dataset curated by other researchers in previous work.
The trained models included:
- ResNet-9
- ResNet-18
- ResNet-34
- ResNet-50
- EfficientNet_B0
The models trained from scratch proved to be more efficient, with ResNet-9 achieving a validation accuracy of 99.37% and a validation loss of 0.020, while the best performing pre-trained model had a validation accuracy of 97.98% and validation loss of 0.096. EfficientNet managed to compete with ResNet, while offering much lower training times, making this architecture a very good solution for real-world applications.