| Title |
Comparative Generative AI Techniques for Effective Fire Blight Leaf Image Generation |
| Authors |
이지민(Ji-Min Lee) ; (RI ZHENG) ; (HELIN YIN) ; (DONG JIN) ; 구영현(Yeong Hyeon Gu) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.6.38 |
| Keywords |
Data augmentation; Fire blight; Generative adversarial networks(GAN); Image-to-image translation; Style transfer |
| Abstract |
Agricultural damage caused by plant disease is increasing every year. With the recent development of artificial intelligence technology, research for diagnosing plant disease using deep learning-based image recognition technology is actively underway. Deep learning models require a large amount of learning data, but it is really difficult to collect plant disease images due to the seasonality and lack of professional workforce. In particular, rare diseases such as Fire Blight are more difficult to collect. Data augmentation is a representative way to solve this problem. Currently, research is being actively conducted to augment the image of pests, but it is mainly augmenting images taken in an artificial environment or diseases images that symptoms of infection appears in the entire leaf area. However, it is necessary to generate realistic images containing various background areas as if they were taken at the farmhouse, not in an artificially created environment, to help improve the performance of the plant disease diagnosis model. In addition, in the case of Fire Blight, the symptoms do not appear in the entire leaf area, but black spots spread along the leaf vein. Therefore, this study explores generative AI techniques to effectively augment the leaf image of Fire Blight. After applying three methods: Traditional (Unconditional) GAN, Image-to-Image Translation, and Style Transfer, the results will be compared and analyzed. |