紀要論文

植物病害自動診断における効果的な全体・茎頂部画像のデータ拡張手法の提案

PROPOSAL OF AN EFFECTIVE DATA AUGMENTATION METHOD FOR WHOLE PLANT AND STEM TOP IMAGES IN AUTOMATED DIAGNOSIS OF PLANT DISEASE

jpn

24368083

https://doi.org/10.15002/00030724

法政大学大学院理工学研究科

法政大学大学院紀要. 理工学研究科編

65

1

5

2024-03-24

Version of Record

In recent years, automatic plant disease diagnosis systems based on deep learning techniques have been proposed. These studies have mainly used images of plant leaf surfaces as input, and there has been no progress in the development of diagnosis systems for other parts of the plant. In addition, while high discrimination accuracy can be obtained for images taken in the same field (shooting environment) as that of the training data, there is still a problem that the discrimination accuracy for images taken in different fields (unknown data) is significantly reduced. Therefore, in this study, we focused on the diagnosis of whole plants and stem tops, and constructed an automatic diagnosis system using a high-performance convolutional neural network model. In addition, a new data expansion method utilizing latent diffusion model is introduced. We trained images of diseases for which data is limited, generated thousands of similar images based on them, and incorporated them into the dataset. In addition, to increase the diversity of the data, we generated healthy images taken in different fields with the disease symptoms of each disease. These images were added to the dataset and the performance of the classification model was verified. The improved generalization performance of the classification model confirms the usefulness of the proposed model.

Article