<b>A study on an efficient citrus Huanglong disease detection algorithm based on three-channel aggregated attention</b>

Addressing the complexity and diversity of field symptoms related to citrus Huanglong disease (HLD), which significantly impacts yield and quality, while also tackling the challenges associated with the low efficiency and inadequate accuracy of traditional detection methods, this study introduces a field symptom detection model based on an enhanced version of YOLOv8. Initially, image enhancement techniques and other preprocessing steps were applied to samples of citrus Huanglong disease collected from various locations in Fujian Province to improve the model's capability to detect HLB in complex environments. Subsequently, a novel C2f Attention IRMB module was developed to replace the original cross-stage local layer convolution C2f, thereby providing a more comprehensive consideration of target position and size differences, improving target localization, and enhancing the model's ability to extract relevant features of citrus Huanglong disease while achieving model lightweighting. Additionally, a three-channel aggregated attention module was integrated into the Neck Powerneck, effectively capturing richer global and local image information, fostering interaction and fusion among different feature layers, and reducing redundancy and repeated computations to enhance the overall efficiency and performance of the model. The model's capacity to express features of citrus Huanglong disease was further strengthened by embedding a finely designed attention mechanism. Lastly, an efficient detection head was constructed to accelerate the inference process. Experimental results demonstrate that on a dataset comprising 12 diseases and 2 healthy states, the mean Average Precision at IoU 0.50 (mAP50) of this model achieved 97%, with a Precision of 91.5%; these figures reflect improvements of 1.1% and 1.2%, respectively, compared to the original YOLOv8 algorithm. Additionally, the inference speed increased by 14.6%, fully meeting the real-time requirements for detecting diseases in citrus fields and illustrating the effectiveness and advanced nature of the improved algorithm, thereby providing robust support for the rapid identification of diseases in the citrus cultivation process.针对显著影响产量与品质的柑橘黄龙病（citrus Huanglong disease, HLD）田间症状的复杂性与多样性，同时解决传统检测方法效率低下、精度不足的难题，本研究提出一种基于改进版YOLOv8的田间症状检测模型。首先，针对从福建省多地采集的柑橘黄龙病样本，采用图像增强等预处理手段，以提升模型在复杂环境下的病害检测能力。随后，设计了一种新型C2f注意力IRMB模块，替换原跨阶段局部卷积C2f模块，该模块可更全面地考量目标位置与尺寸差异，优化目标定位能力，提升模型对柑橘黄龙病相关特征的提取能力，同时实现模型轻量化。此外，在特征融合颈部（Neck）的Powerneck结构中集成了三通道聚合注意力模块，可有效捕获更丰富的图像全局与局部信息，促进不同特征层间的交互与融合，减少冗余与重复计算，从而提升模型整体效率与性能。通过嵌入精心设计的注意力机制，进一步强化了模型对柑橘黄龙病特征的表达能力。最后，构建了高效检测头以加快推理速度。实验结果表明，在包含12种病害与2种健康状态的数据集上，该模型在交并比（Intersection over Union, IoU）为0.50时的平均精度均值（mean Average Precision, mAP50）达到97%，精确率（Precision）为91.5%；与原始YOLOv8算法相比，这两项指标分别提升了1.1%与1.2%。此外，推理速度提升了14.6%，完全满足柑橘田间病害检测的实时性要求，验证了改进算法的有效性与先进性，可为柑橘种植过程中的病害快速识别提供有力支撑。