摘要
为实现林间荔枝快速准确检测,本研究提出了一种基于YOLOv4-Tiny改进的荔枝轻量化检测方法.通过在主干网络加入SPP模块和在颈部网络加入ECA模块,提升模型对小目标荔枝的识别效果和在复杂背景下的检测性能,消融实验验证了改进方法的有效性.利用林间荔枝图像数据集训练了改进的模型,并分析了改进前后的性能差异,结果显示,改进后的模型检测速度为54帧/秒,精确率、召回率、平均精度分别为99.20％、82.88％、95.49％.与SSD、Faster RCNN及 YOLOv4-Tiny模型相比,改进后的模型平均精度提升了 15.23％、17.33％、5.57％,召回率提升了 10.98％、11.52％、7.22％.本研究可为荔枝的生长监测、机械采摘和人工估产等提供技术支持.
Abstract
To achieve fast and accurate detection of lychee in the forest,this study proposes a lightweight detection method for lychee based on YOLOv4-Tiny improvement.By adding SPP module to the backbone network and ECA module to the neck network,the model improves the recognition effect of small-target lychee and the detection performance in complex backgrounds,and the ablation experiments verify the effectiveness of the improved method.The improved model was trained using the forest lychee image dataset,and the performance differences before and after the improvement were analyzed,and the results show that the improved model detects as fast as 54 frames/second,and the precision,recall,and average accuracy are 99.20％,82.88％,and 95.49％,respectively.Compared with SSD,Faster RCNN,and YOLOv4-Tiny model,the improved model boosted the average precision by 15.23％,17.33％,and 5.57％respectively,and the recall by 10.98％,11.52％,and 7.22％respectively.This study can provide technical support for research related to growth monitoring,mechanical harvesting and artificial yield estimation of lychee fruit.
NETL
NSTL
万方数据