Note (General)type:Thesis
With the breakthrough of deep learning techniques, many excellent applications for the automated diagnosis of plant disease have been proposed. However, there are several open issues for developing practical plant disease diagnosis systems in real cultivation. Firstly, most conventional methodologies only accept narrow range images, typically one or quite a limited number of targets are in their inputs. Applying these models to wide-angle images in large farms would be very time-consuming, since many targets (e.g., leaves) need to be diagnosed. In this work, we propose a two-stage system which has independent leaf detection and leaf diagnosis stages for wide-angle disease diagnosis. We show that our proposal attains a promising disease diagnostic performance that is more than six times higher than end-to-end systems (state-of-the-art detection methods like Faster R-CNN or SSD) with F1-score of 33.4 - 38.9% compared to 4.4 - 6.2% on an unseen target dataset.Secondly, the lack of image resolution (i.e., diagnosing from low-quality input images such as low-resolution, blur, poor camera focus, etc.) could significantly reduce the diagnostic performance in practice. Also, high-resolution data is very difficult to obtain and are not always available in practice. Deep learning-based techniques, and particularly generative adversarial networks (GANs), can be applied to generate high-quality super-resolution images, but these methods often produce unexpected artifacts that can lower the diagnostic performance. In this paper, we propose a novel artifact-suppression super-resolution method that is specifically designed for diagnosing leaf disease, called LASSR. Our LASSR can detect and suppress artifacts to a considerable extent. Thus, generating much more pleasing, high-quality images from low-resolution inputs. Experiments show that training with data generated by our proposal significantly boosts the performance on an unseen test dataset by over 21% compared with the baseline.
Thirdly, collecting and labeling training disease data for these diagnosis systems requires solid biological knowledge and is very labor-intensive. Limited amount of disease training data leads to the fourth problem of model overfitting. The performance of disease diagnostic models are drastically decreased when used on test data sets from new environments. Meanwhile, we observe that healthy images are easier to collect. Based on this, we propose LeafGAN, a novel image-to-image translation system. LeafGAN generates countless diverse and high-quality diseased data via transformation from healthy images, as a data augmentation tool for improving the performance of plant disease diagnosis. Our model can transform only relevant areas from images with a variety of backgrounds, thus enriching the versatility of the training images. Experiments show that data augmentation with LeafGAN help to improve the generalization, boosting the diagnostic performance on unseen data by 7.4% from baseline.In summary, we show that our approaches significantly improve the diagnostic performance under practical settings, confirming to be efficient and reliable methods for real cultivation scenarios.
Collection (particular)国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文
Date Accepted (W3CDTF)2022-01-10T16:22:37+09:00
Data Provider (Database)国立国会図書館 : 国立国会図書館デジタルコレクション