| 鲁远,胡威.基于深度学习的柑橘病害智能检测技术研究进展[J].中国南方果树,2026,55(4):301-310 |
| 基于深度学习的柑橘病害智能检测技术研究进展 |
| Research Progress on Intelligent Detection Technology of Citrus Diseases Based on Deep Learning |
| 投稿时间：2026-02-14 修订日期：2026-04-24 |
| DOI： |
| 中文关键词: 柑橘病害 深度学习 卷积神经网络 智能检测 图像识别 |
| 英文关键词:Scitrus disease deep learning convolutional neural network (CNN) intelligent detection image recognition |
| 基金项目:江西省自然科学基金（20242BAB23064，20252BAC240596）；国家自然科学基金项目（32360666） |
|
| 摘要点击次数: 9 |
| 全文下载次数: 0 |
| 中文摘要: |
| 柑橘是我国第一大水果作物,病害防控是保障产业可持续发展的关键。针对传统人工识别方法效率低、成本高、主观性强等局限,本文系统梳理了深度学习技术在柑橘病害智能检测领域的最新研究进展。文章首先概述了柑橘黄龙病、溃疡病、炭疽病等主要病害的症状特征与危害程度,分析了基于视觉特征进行智能识别的可行性；随后对比了传统机器学习与深度学习方法在特征提取、识别精度、泛化能力等方面的差异；重点综述了卷积神经网络(CNN)及其改进模型(AlexNet、VGGNet、ResNet、MobileNet)在柑橘病害识别中的应用,详细分析了不同网络架构的性能特点；针对移动端部署需求,探讨了轻量化模型设计策略；结合 耘眼 等实际应用案例,总结了深度学习模型在田间环境下的部署经验。研究表明,深度学习模型对柑橘病害的识别准确率普遍超过90%,部分模型达到98%以上,显著优于传统方法。但实际应用仍面临高质量数据集构建困难、复杂田间环境适应性不足、模型可解释性弱、实时检测能力有限等挑战。未来应重点突破小样本学习、多模态数据融合、早期病害检测、轻量化模型优化等关键技术,推动深度学习技术从实验室走向生产一线,为柑橘产业智能化发展提供技术支撑。 |
| 英文摘要: |
| Citrus is the largest fruit crop in China, and disease control is crucial for ensuring the sustainable development of the industry. To address the limitations of traditional manual identification methods, including low efficiency, high cost, and strong subjectivity, this paper systematically reviews the latest research progress of deep learning technology in the intelligent detection of citrus diseases. The article first outlines the symptom characteristics and damage severity of major citrus diseases such as Huanglongbing (citrus greening), canker, and anthracnose, and analyzes the feasibility of intelligent identification based on visual features. Subsequently, it compares the differences between traditional machine learning and deep learning methods in terms of feature extraction, recognition accuracy, and generalization capability. The review focuses on the application of Convolutional Neural Networks (CNN) and their improved models (AlexNet, VGGNet, ResNet, MobileNet) in citrus disease identification, with detailed analysis of the performance characteristics of different network architectures. In response to mobile deployment requirements, lightweight model design strategies are discussed. Combined with practical application cases such as "Yunyan," the paper summarizes deployment experiences of deep learning models in field environments. Research indicates that deep learning models generally achieve over 90% accuracy in citrus disease identification, with some models reaching above 98%, significantly outperforming traditional methods. However, practical applications still face challenges including difficulties in constructing high-quality datasets, insufficient adaptability to complex field environments, weak model interpretability, and limited real-time detection capabilities. Future efforts should focus on breakthroughs in key technologies such as few-shot learning, multi-modal data fusion, early disease detection, and lightweight model optimization, to promote the transition of deep learning technology from laboratory to production frontlines and provide technical support for the intelligent development of the citrus industry. |
| 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
