| 唐卫东,陈冠华,王瑞,刘秋明,刘振文.基于GAN与ResNet34-UNet的黄瓜叶片病斑分割方法[J].井冈山大学自然版,2025,46(2):68-80 |
| 基于GAN与ResNet34-UNet的黄瓜叶片病斑分割方法 |
| CUCUMBER LEAF SPOT SEGMENTATION METHOD BASED ON GAN AND ResNet34-Unet |
| 投稿时间：2024-10-11 修订日期：2024-12-16 |
| DOI：10.3969/j.issn.1674-8085.2025.02.009 |
| 中文关键词: 叶片病斑分割 GAN U-Net 深度学习 注意力机制 |
| 英文关键词: leaf lesion segmentation GAN U-Net deep learning attention mechanism |
| 基金项目:国家自然科学基金项目(31860574)；江西省自然科学基金项目(20224BAB205025) |
| 作者 | 单位 | E-mail | | 唐卫东 | 江西理工大学软件工程学院, 江西, 南昌 330013
井冈山大学电子与信息工程学院, 江西, 吉安 343009 | metangwd@163.com | | 陈冠华 | 江西理工大学软件工程学院, 江西, 南昌 330013
井冈山大学电子与信息工程学院, 江西, 吉安 343009 | | | 王瑞 | 武警江西省总队吉安支队, 江西, 吉安 343009 | | | 刘秋明 | 江西理工大学软件工程学院, 江西, 南昌 330013 | | | 刘振文 | 井冈山大学电子与信息工程学院, 江西, 吉安 343009 | |
|
| 摘要点击次数: 92 |
| 全文下载次数: 222 |
| 中文摘要: |
| 针对现有叶片病斑分割方法存在病斑分割不完整与小区域病斑分割困难等问题，本研究提出一种基于改进 GAN 的叶片病斑分割方法，旨在通过智能化病斑分割来提高作物病害防治能力。该模型由生成器与判别器构成，其中，生成器网络 ResNet34-UNet 是在 U-Net 基础上，结合改进的 ResNet34 和注意力机制组成，主要用于生成更逼真的分割结果以欺骗判别器，而判别器网络则是由一个深度卷积神经网络构成，用于区分生成的分割结果与真实标签。实验结果表明，该方法能够较好地实现叶片病斑分割，其 Sensitivity、Specificity、Dice、Accuracy评价指标分别达到 90.36%、97.72%、85.25%和 97.10%，与其他分割网络相比，该方法保留了更多细节信息，对小区域病斑也能实现较为完整的分割，可为黄瓜病害识别与防治管理提供支撑。 |
| 英文摘要: |
| In view of the problems of incomplete segmentation and difficult segmentation of small area lesions in the existing segmentation methods for leaf lesions, a leaf spot segmentation method based on improved Generative Adversarial Network (GAN) was introduced, focusing on enhancing the capability of crop disease prevention and control by intelligent spot segmentation. This model consisted of a generator and a discriminator. The generator network, ResNet34-UNet, consisted of U-Net and improved ResNet34, attention mechanisms. It was primarily used to generate more realistic segmentation results to deceive the discriminator, and the discriminator network was composed of a deep convolutional neural network, tasked with distinguishing between the generated segmentation results and the real labels. The experiments showed that the method could achieve good segmentation of leaf lesions, the evaluation metrics as sensitivity, specificity, dice, and accuracy reached 90.36%, 97.72%, 85.25%, and 97.10% respectively. Compared with the other segmentation networks, the proposed method retained more detailed information, and could achieve more complete segmentation for small areas of disease spots. It could provide strong support for cucumber disease identification and prevention management. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
