Abstract: To address the challenges of strong background interference, significant noise and small lesion targets in red kidney bean leaf disease detection under complex field conditions, this study first constructed a self-built dataset of red kidney bean leaf diseases. Building upon this foundation, a disease detection model named YOLOv11-Mamba is proposed. This model integrates Mamba-based feature extraction with an improved YOLOv11 framework. The core innovation of this model lies in the deep fusion of the advantages of convolutional neural networks (CNN) and state space models (SSM). Specifically, CNNs are employed to effectively extract local detailed texture features of the lesions, while Mamba leverages its exceptional capability in long-sequence modeling to capture the global contextual dependencies in leaf images. This synergistic mechanism significantly enhances the model’s ability to perceive and discriminate subtle lesion features in complex scenarios. Experimental results demonstrate that on a dataset containing four disease types including rust, pest damage, powdery mildew and bacterial leaf spot, YOLOv11-Mamba achieves an 8.5 percentage point improvement, approximately a 24% relative increase in disease detection accuracy compared to the YOLOv11 detection model. Furthermore, while maintaining high inference speed to meet real-time requirements, its performance surpasses that of YOLOv11 segmentation model by 18.5%. This research provides a lightweight solution with promising potential for the accurate and efficient visual recognition of agricultural diseases.

摘要： 针对复杂田间环境下红芸豆叶片病害检测面临的背景干扰强、噪声显著以及病斑目标小等挑战，本文首先扩建了一个红芸豆叶片病害数据集，并在此基础上提出一种基于Mamba特征提取与改进YOLOv11的病害检测模型YOLOv11-Mamba。该模型的核心创新在于深度融合了卷积神经网络（Convolutional Neural Network, CNN）与状态空间模型（Structured State Space Model, SSM）的优势。通过CNN有效提取病斑的局部细节纹理特征，同时利用Mamba在长序列建模方面的卓越能力，对叶片图像的全局上下文依赖关系进行建模，这种协同机制显著增强了模型在复杂场景下对细微病斑特征的感知与判别能力。实验结果表明：在锈病、虫害、白粉病和细菌性斑点病四类病害数据上，YOLOv11-Mamba相较于YOLOv11检测模型，病害检测精度提升了8.5个百分点（相对提升约24%）；同时，在保持较高推理速度、满足实时性要求的前提下，其性能较YOLOv11分割模型提升了18.5%。本文为农业病害的精准、高效视觉识别提供了一种具备良好推广前景的轻量化解决方案。