基于改进YOLOv8s的葡萄果实与果梗检测方法

打开文本图片集

中图分类号：S663.1；TP391.4 文献标识码：A 文章编号：2095-5553（2025）12-0103-11

Abstract： Accurate detection of grape fruits and stalks was the key research content to realize inteligent grape harvesting. Theacuracyandreal-time performanceofthecurrent detectionalgorithms didnotmettheneedsofthe actual production environment，and it was dificult torealize theeficient detectionofmultiple bunchesof grapes incomplexscenarios.To address these problems，this paper proposed a high-precision real-time method calledRepViT—YOLOv8s basedon YOLOv8s，which replaced theconvolutional block of the baseline model with RepViT in the image featureextraction stage.Bycombining thewidereceptivefieldof theVision Transformerwiththeeficiencyofconvolutional neural networks， RepViT enabled richer image feature extraction and finer-grained target representation. It also proposed a C2f— CBAMmodule in the feature fusion stage，where theCBAMvisual atention mechanism highlightedtarget featuresand suppressed background information，while the C2f module enhanced targetfeature extraction.For bounding boxregression， SIoULoss was employedto more acuratelycomputedeviations between predictedand ground-truth boxes，leading to faster modelconvergence during training.Finally numerous experimentalresultsshowed that comparedwith YOLOv8s，RepViT—YOLOv8s achieved improvements of 3.5% ， 2.5% ， 3.0% and 2.3% in P ，R， mAP@0.5 ，and mAP@0.5：0.95 respectively，on the MS COCO 2O17 validation set.Moreover，on a self-constructedgrape fruit and stalks dataset， RepViT-YOLOv8sachieved 97.4% precision， 79.6% recall， 86.8% mAP@0.5 and 61.6% ， outperforming YOLOv8s by 3.3% ， 3.0% ， 3.2% and 2.6% ，respectively.

Keywords： grape fruit detection； grape stalk;deep learning; object detection;convolutional neural network

0 引言

葡萄果实口感鲜美、富含多种营养元素，副产品种类多样、经济价值高，因此葡萄种植面积和产量逐年攀升1。（剩余17431字）