基于改进YOLOv8n的草莓叶片病害检测方法

Strawberry leaf disease detection method based on improved YOLOv8n

XIA Shunxing， NI Ming， LUO Youlu， HE Yinghao， ZHAO Taotao（Collegeof InformationEngineering，SichuanAgricultural University，Ya'an625O14，China）

Abstract：Inordertoimprovethedetectionabilityof targetdetectionmodelsforstrawberryleaf diseasesinorchard environment，thisstudyusedtheYOLOv8nmodelasthebaseline model，introducedthedynamicconvolution module toreplacethethirdconvolution layerof the backbone network and partoftheC2fmoduleof theneck network，introduced the GSConv and Slim-neck module toreplacetheconvolution layerand partof the C2f moduleof the neck network，and introduced the content-aware reassemblyof features （CARAFE）operator to replacethe nearestneighbor interpolation method in upsampling.An improved YOLOv8n model named YOLOv8n-DGC was proposed to improve thedetection accuracy of strawberyleaf diseases while maintaining the lightweightofthe model.Theresults showed that the meanaverage precisionwhen the intersection over union（ IoU ）threshold was 0.50 （ mAP₅₀ ），the mean average precision when the IoU was between （20 0.50 and 0.95 （ mAP_50：95 ），precision and recall rate of the improved model YOLOv8n-DGC for strawberry leaf disease detection were 2.5percentage points，1.5percentage points，.6percentage pointsand1.6percentagepoints higher thanthose of the baseline model，respectively. The model size and parameter quantity increased by 3.2% and 3.3% ，respectively， while the number of floating point of operations decreased by 8.6% . Compared with models such as Faster R-CNN，SSD， YOLOv5s，and YOLOv7-tiny，the YOLOv8n-DGC model beter achieved a balance between detectionaccuracyand ffi

Keywords：strawberry；leaf diseases；object detection；YOLOv8n；dynamic convolution;GSConv；CARAFE

草莓属蔷薇科浆果类多年生草本植物，素有“水果皇后”的美称。草莓果实颜色艳丽，芳香多汁，且富含糖类、氨基酸、维生素以及钙、磷、铁等多种营养物质[1-2]，具有抗氧化、抗癌、抗炎等功能[3-4]。近年来，中国草莓种植面积均较大，产量也较高[5]。草莓栽培过程中，病害是导致草莓产量减少、品质下降的重要原因之一，及时并准确识别病害类型对草莓安全生产、果农效益提高具有重要意义。

近年来，随着计算机技术和人工智能的发展，特别是深度学习技术的应用，作物病害检测技术得到了显著的提升。深度学习是一种基于人工神经网络的机器学习方法，具有强大的特征提取和模式识别能力[]通过对大量数据的学习，深度学习模型能够自动提取图像中的复杂特征，实现高效、准确的病害检测。目前目标检测领域的深度学习方法主要分为两类：一类是以R-CNN[7]、FastR-CNN[8]、FasterR-CNN[9]为代表的两阶段检测算法，该类算法分两步完成，首先生成候选目标区域，然后对候选目标区域进行分类和定位：另一类是以YOLO系列算法[10-13] SSD 算法[14]为代表的单阶段检测算法，该类算法同时进行边界框位置判断和类别分类。相比较而言，单阶段分类算法简化了处理流程，因而具有更快的检测速度。

由于YOLO算法在特征提取、识别实时性、泛化性等方面的优势，目前，YOLO算法在作物病虫害识别及果实目标检测等领域得到了广泛应用。杨宇游等[15]在YOLOv5模型的主干网络中引入BoTNet模块，并用GIoU-NMS算法替换原来的非极大值抑制（NMS）算法，对YOLOv5模型进行优化，改进后的模型对草莓病害的识别精确率和平均精度均值分别比原模型提升2.1个百分点和1.2个百分点。……