基于改进YOLOv8s的轻量化冬枣识别方法

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2095-5553（2026）04-0126-08

Abstract：To solve the problemof rapidrecognition of winterjujube innatural environment and limitedcomputing resources for edge devices，alightweight recognition method of winter jujube，YOLOv8s—SFl，basedon YOLOv8s model was proposed，soas to deploy the model toedge devices later.Firstly，the backbone extraction network of YOLOv8s is replacedby StarNetlightweight network toreduce thecomplexityof theoriginalmodel.Secondly，FasterNetis introduced intoC2f moduleto buildanew structure，C2f—Faster，toreduce theconsumptionofcomputingresources.Finally，anew networkdetection head，Detect_LSCD，isproposed to further reduce the storage capacityand modelcomplexity.The experimental results show that the frame rate of YOLOv8s—SFl model is 313.5 frames/s，which is 14.66% higher than thatof the original model，while the weight size and parameter numberare 6.8MB and 3.471M ，respectively，which are reduced by 68.37% and 68.80% compared with the original model. The Precision rate，recallrate and average accuracy areonlyreduced by about 1% compared with the original model. The model size of the improved YOLOv8s—SFl is reduced by 61.58% ， 68.37% ， 53.10% and 56.96% compared to the YOLOv5s，YOLOv8s，YOLOv9sand YOLOvlOs，respectively. The number of parameters decreased by 61.90% ， 68.80% ， 51.57% and 51.91% （204 respectively，and the floating-point calculation amount decreased by 50.42% ， 58.45% ， 55.80% and 44.86% ：

Keywords：winter jujube；natural environment；lightweight models；YOLOv8s；inteligent harvesting

0 引言

冬枣的人工采收严重制约该产业规模化发展[1]。（剩余11938字）