ようこそ ゲスト さん
ログイン
入力補助
English
カテゴリ
インデックスツリー
ランキング
アクセスランキング
ダウンロードランキング
その他
法政大学
法政大学図書館
インデックスツリー
資料タイプ別
学内論文
大学院紀要=Bulletin of graduate studies
法政大学大学院紀要. 理工学・工学研究科編
法政大学大学院紀要. 理工学研究科編
法政大学大学院紀要. デザイン工学研究科編
法政大学大学院紀要. 情報科学研究科編
法政大学懸賞論文優秀論文集
法政大学国際文化学部国際社会演習トランスナショナル・ヒストリー研究卒業論文集
このアイテムのアクセス数:
4
件
(
2024-11-02
08:39 集計
)
Permalink : https://doi.org/10.15002/00030752
Permalink : https://hdl.handle.net/10114/00030752
閲覧可能ファイル
ファイル
フォーマット
サイズ
閲覧回数
説明
gradse_65_22R4126
pdf
413 KB
28
論文情報
ファイル出力
アイテムタイプ
紀要論文
タイトル
植物病害診断システムにおけるDiffusion-Basedデータ拡張と補完情報の効果
その他のタイトル
EFFECTS OF DIFFUSION-BASED DATA EXTENSION AND COMPLEMENTARY INFORMATION ON PLANT DISEASE DIAGNOSIS SYSTEMS
著者
著者名
西井, 彩葉
著者名
NISHII, SAYO
言語
jpn
ISSN
24368083
DOI
https://doi.org/10.15002/00030752
出版者
法政大学大学院理工学研究科
雑誌名
法政大学大学院紀要. 理工学研究科編
巻
65
開始ページ
1
終了ページ
7
発行年
2024-03-24
著者版フラグ
Version of Record
キーワード
plant disease diagnosis
Stable diffution
DreamBooth
抄録
Plant diseases are a serious problem in agriculture, with the FAO reporting that up to 40% of the world’s annual crop production is lost to pests. This causes losses to the global economy in excess of $220 billion annually, necessitating rapid detection and accurate treatment of plant diseases. However, current inspection methods are mainly visual and genetic, which are costly in terms of manpower, money, and time. To address this issue, research has developed inexpensive and convenient automatic plant disease diagnosis systems. 2015 saw the development of an automatic diagnosis system using deep learning, followed by the release of plant disease datasets and the proposal of improved versions, which have achieved high discrimination performance. However, these automatic diagnosis systems have the problem that the leaf images included in the training images are pre-cropped with the same background, which significantly reduces the diagnosis performance for images taken in actual fields or agricultural fields. As a means of addressing this problem, it is essential to ensure the diversity of training images and improve upon realistic datasets. Many data extensions using GAN-based image generation models have been proposed in the past, but these methods also suffer from a lack of diversity due to a biased distribution of training images of healthy leaves, as well as from the enormous amount of time and effort required. On the other hand, research on Diffusion Models (DMs) has been progressing, and it is now possible to generate images with rich diversity using large-scale pre-trained models. In this study, we applied DreamBooth to Stable diffusion to generate disease-specific images using a small number of plant disease images, and evaluated the effectiveness of Diffusion-based data expansion using the generated images. The results showed that the +Prompt(v2)-augmentation method achieved the best discrimination performance for cucumber leaves in HE, MYSV, MD, Macro avg, and Micro avg, and for eggplant leaves in HE, PM, and BW. This means that we can use the information complemented by prompt to identify the leaves. We confirmed that domain-adaptation is also feasible by considering complementary information from prompts. Future work is expected to improve the quality of the generated images by addressing problems in converting latent representations to images in the text-to-image generation model, adjusting the loss function of DreamBooth, and improving the attention mechanism of U-Net. Translated with DeepL.com (free version)
資源タイプ
Article
インデックス
資料タイプ別
 > 
学内論文
 > 
法政大学大学院紀要. 理工学研究科編
501 学内論文
 > 
紀要
 > 
法政大学大学院紀要. 理工学研究科編
 > 
65
ホームへ戻る