摘要
针对自然场景下苹果叶片病害检测存在背景噪声、病害表征相似及小尺度目标多等因素导致模型检测精度不足的挑战,提出一种ALD—YOLO算法,旨在实现自然场景下苹果叶片病害的精准检测。首先,提出一种CBAM—G注意力机制并嵌入主干网络,使网络在特征提取阶段聚焦病害关键特征并降低图片形变产生的影响。其次,将多层次特征信息输入OCR模块并反馈给头部网络,使模型获取更为丰富的像素级语义信息从而精准分类。最后,提出一种动态赋权IoU损失函数,通过为小尺度目标动态赋权来提高小尺度目标损失,进而提升对小尺度目标的检测精度。结果表明,该算法在自制数据集上的精确率为84.93%、召回率为72.88%、平均精度均值为77.48%、处理速度达24.74帧/s。
To address the challenges in apple leaf disease detection in natural scenes,such as background noise,similar disease symptoms,and a high density of small-scale targets,this study proposes the ALD—YOLO algorithm to achieve accurate detection of apple leaf diseases in natural environments.First,a CBAM—G attention mechanism is proposed and integrated into the backbone network,enabling the network to focus on key disease features during feature extraction while reducing the impact of image distortions.Second,multi-level feature information is fed into an OCR module and then returned to the head network,providing the model with richer pixel-level semantic information for precise classification.Finally,a dynamic weighted IoU loss function is proposed,dynamically assigning weights to small-scale targets to increase their loss,thereby improving detection accuracy for such targets.Experimental results demonstrate that the algorithm achieves an accuracy of 84.93%,a recall rate of 72.88%,and a mean average precision(mAP)of 77.48%on a custom dataset,with a processing speed of 24.74 frames per second.
作者
刘霞
周家横
古庆辉
Liu Xia;Zhou Jiaheng;Gu Qinghui(Changjiang Polytechnic,Wuhan,430074,China;College of Horticulture&Forestry Science,Huazhong Agricultural University,Wuhan,430070,China;School of Electrical Engineering and Automation,Wuhan University,Wuhan,430072,China)
出处
《中国农机化学报》
北大核心
2026年第1期52-61,共10页
Journal of Chinese Agricultural Mechanization
基金
国家自然科学基金(31801258)
2023年度湖北省教育厅科学研究计划指导性项目(B2023612)。
