Study on Lightweight Apple Detection Model for Smart Orchards

HU Junfeng; LIU Zilong; LIU Dayang

doi:10.15933/j.cnki.1004-3268.2026.04.014

Abstract：In the development of automated agricultural harvesting equipment,high-precision recognition,real-time response,and lightweight design are core requirements for target detection algorithms.To address these technical challenges,this study innovatively proposed a lightweight apple detection model named YEMB-FPN(YOLOv10n-efficient multi-branch FPN).The model replaced the last two C2f modules in the backbone network of the original YOLOv10n with the grouped efficient multi-scale(GEMS)convolution module.In the neck network,a bidirectional feature pyramid framework was adopted,where the original Concat operation was replaced by a weighted bidirectional feature pyramid fusion module(BiFPN_Fusion).Additionally,a novel cross-stage multi-scale fusion block(CS-MSFB)was designed to substitute all original C2f and C2fCIB modules in the neck network.The model was validated using a self-constructed apple target detection dataset from orchard environments.Experimental results demonstrated that the YEMB-FPN model achieved an mAP50 of 98.6％,with a model size of only 4.6 MB.Compared to the baseline YOLOv10n,this represented a 3.3 percentage point improvement in mAP50 and a reduction of 1.2 MB in model size.These findings indicated that the proposed algorithm significantly enhanced apple detection accuracy,while its lightweight design markedly improved compatibility with low-computational hardware platforms,which provided critical technical support for the embedded deployment of intelligent agricultural equipment.