摘要
为实现自然环境下不同成熟度火龙果在不同光照、不同遮挡场景下的精确快速识别,提出了一种基于对YOLOv5的网络模型改进的一种检测模型(PITAYA-YOLOv5)。首先,使用k-means++算法重新生成火龙果数据集的锚框,提高了网络的特征提取能力;其次,将CSPDarkNet替换成PPLCNet作为骨干网络,并加入SE注意力模块(Squeeze-and-Excitation block),在降低网络参数量的同时保持检测精度;同时加入加权双向特征金字塔网络(Bi-FPN)替换YOLOv5的特征融合网络,提高网络对不同尺度特征图的融合效率;引入αDIoU损失函数,提高了模型的收敛效果。试验结果表明:PITAYA-YOLOv5目标检测模型的平均精度均值为94.90%,较原模型提高1.33个百分点,F1值为91.37%,较原模型提高1.12个百分点,平均检测速度达到20.2 ms,占用内存仅有8.1 M。针对枝条遮挡和果间遮挡下的火龙果检测能力明显增强。对比Faster R-CNN、CenterNet、YOLOv3、YOLOv5以及轻量化骨干网络ShuffleNetv2,该模型具有良好的检测精度和实时性。该模型能够有效地在自然环境下识别并检测火龙果的成熟度,为果园智能管理与早期产量预估等提供了参考。
The accurate identification of pitaya maturity is very important for intelligent management and early yield estimation.In order to realize the accurate and fast recognition of maturity of pitaya under different illumination and different occlusion in the natural environment,this study proposed a recognition method(PITAYA-YOLOv5)based on the network model of YOLOv5.Firstly,k-means++algorithm was used to regenerate the anchor frame of Pitaya dataset,which improved the feature extraction ability of the network.Secondly,the CSPDarknet was replaced by PPLCNet as backbone network,and SE attention block(Squeeze-and-Excitation block)was added to reduce the amount of network parameters while maintained the detection accuracy.At the same time,a weighted Bi-directional Feature Pyramid network was added to replace the feature fusion network of YOLOv5 to improve the fusion efficiency of the network for feature maps of different scale.TheαDIoU loss function was introduced to improve the convergence effect of the model.The results of ablation test showed that the mean average precision of the PITAYA-YOLOv5 target detection model is 94.90%,1.33 percentage points higher than that of the original model;F1 score was 91.37%,1.12 percentage points higher than that of the original model;the average detection speed was up to 12.2 ms,and the memory occupied was only 8.1 M.The detection ability of pitaya under branch occlusion and inter fruit occlusion was significantly enhanced.Compared with Faster R-CNN,CenterNet,YOLOv3,YOLOv5,and the lightweight backbone network ShuffleNetv2,the model has good detection accuracy and real-time performance.This model can effectively identify and detect the maturity of pitaya fruit in natural environments.It may provide reference for the intelligent management of orchard and early yield estimation.
作者
马瑞峻
何浣冬
陈瑜
赖宇豪
焦锐
唐昊
MA Rui-jun;HE Huan-dong;CHEN Yu;LAI Yu-hao;JIAO Rui;TANG Hao(College of Engineering,South China Agricultural University,Guangzhou 510642,China)
出处
《沈阳农业大学学报》
CAS
CSCD
北大核心
2023年第2期196-206,共11页
Journal of Shenyang Agricultural University
基金
广东省科技计划项目(2021B1212040009)。
