紀要論文

植物病害自動診断システムにおけるドメイン適応の試み

AN ATTEMPT AT DOMAIN ADAPTATION IN AUTOMATIC PLANT DISEASE DIAGNOSIS SYSTEMS

jpn

24368083

https://doi.org/10.15002/00025385

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

63

1

5

2022-03-24

Version of Record

In the automatic diagnosis system of plant diseases, there is a problem that the classification performance of the data obtained in the same field as the training data is high, but that of the data in the field not used for training is greatly reduced. The reason for this is that images taken in the same environment often have similar backgrounds, disease appearance, etc., and the same photographic equipment is often used, so the system overfitting on information specific to these fields. In recent years, the field of domain adaptation has been studied in order to maintain the performance on such different training data and data obtained from different distributions. In this paper, we implemented one unsupervised domain adaptation method and two domain standardization methods in a practical setting for the task of automatic plant disease diagnosis, and compared and verified the results. The results showed that the domain adaptation method improved the performance of the task by a small amount, but did not lead to a significant improvement. We also checked whether the current automatic plant disease diagnosis system has the information of the field where the images were taken, and investigated how much the current model depends on the information of the field. From the experiments, it was confirmed that the current disease classification model has a lot of field information and may be using the field information for disease classification.

Article