Do different weight initialization strategies have an impact on transfer learning for plant disease detection?

Abstract

Abstract The concept of weight initialization technique for transfer learning refers to the practice of using pre-trained models that can be modified to solve new problems, instead of starting the training process from scratch. By using pre-trained models as a starting point, the network can learn from patterns and features present in the original data, improving overall accuracy and allowing for faster convergence during training. In this study, four different transfer learning weight initialization strategies are proposed for plant disease detection: random initialization, pre-trained model on different domain (ImageNet), model trained on related domain (ISIC 2019), and model trained on same domain (PlantVillage). Weights from each strategy are transferred to a target dataset, Plant Pathology 2021. These strategies were implemented using four state-of-the-art CNN-based architectures: AlexNet, DenseNet, MobileNetV2, and VGG. The best result was obtained when both the target and source datasets included images of plant diseases. In this case, VGG was used and resulted in an 85.9% weighted f-score, which is a 9% improvement from random initialization. The transfer of knowledge from small-sized, related domain data (skin cancer data) was almost as successful as the transfer from ImageNet. Transferring from ImageNet yielded an f-score of 85.7%, while transferring from skin cancer data resulted in an f-score of 85.2%. This indicates that ImageNet, which is widely favored in the literature, may not necessarily represent the most optimal transfer source for the given context. Finally, the classifications made by the proposed models were visualized using Grad-CAM to better understand the decision-making process.