In this study,we propose Space-to-Depth and You Only Look Once Version 7(SPD-YOLOv7),an accurate and efficient method for detecting pests inmaize crops,addressing challenges such as small pest sizes,blurred images,low...In this study,we propose Space-to-Depth and You Only Look Once Version 7(SPD-YOLOv7),an accurate and efficient method for detecting pests inmaize crops,addressing challenges such as small pest sizes,blurred images,low resolution,and significant species variation across different growth stages.To improve the model’s ability to generalize and its robustness,we incorporate target background analysis,data augmentation,and processing techniques like Gaussian noise and brightness adjustment.In target detection,increasing the depth of the neural network can lead to the loss of small target information.To overcome this,we introduce the Space-to-Depth Convolution(SPD-Conv)module into the SPD-YOLOv7 framework,replacing certain convolutional layers in the traditional system backbone and head network.This modification helps retain small target features and location information.Additionally,the Efficient Layer Aggregation Network-Wide(ELAN-W)module is combined with the Convolutional Block Attention Module(CBAM)attention mechanism to extract more efficient features.Experimental results show that the enhanced YOLOv7 model achieves an accuracy of 98.38%,with an average accuracy of 99.4%,outperforming the original YOLOv7 model.These improvements represent an increase of 2.46%in accuracy and 3.19%in average accuracy.The results indicate that the enhanced YOLOv7 model is more efficient and real-time,offering valuable insights for maize pest control.展开更多
目的YOLOv7-tiny(you only look once version 7-tiny)成为实时目标检测领域的常用方法,由于其轻量化网络架构设计和较少的参数量,整个训练过程在单个网络中进行,检测速度快且不需要使用滑动窗口或候选区域,在资源受限、实时性要求高的...目的YOLOv7-tiny(you only look once version 7-tiny)成为实时目标检测领域的常用方法,由于其轻量化网络架构设计和较少的参数量,整个训练过程在单个网络中进行,检测速度快且不需要使用滑动窗口或候选区域,在资源受限、实时性要求高的任务中表现优异。然而,YOLOv7-tiny在特征融合阶段存在相邻层特征融合时信息丢失和非相邻层特征信息差异两个问题。为了解决上述问题,提出一种长短程依赖特征金字塔网络LSRD-FPN(long short range dependency feature pyramid network),并基于该网络对YOLOv7-tiny方法进行改进。方法LSRD-FPN包括两个关键组成部分:局部短程依赖机制SRD(short range dependency)和全局长程依赖机制LRD(long range dependency)。局部短程依赖机制通过改进上采样方式和引入注意力机制,有效缓解了特征融合过程中信息丢失的问题;全局长程依赖机制通过引入跨层连接模块,将主干网络的多尺度特征缩放、融合并分配到检测阶段的不同层级特征。LSRD-FPN不仅增强了模型的特征表达能力,而且提升了其在多尺度目标检测任务的性能表现。结果选用两个不同场景和规模的数据集进行实验。实验结果表明,相较于YOLOv7-tiny,本文方法的mAP分别取得1.3%和0.5%的性能提升。与参数量相当的YOLOv5-s和YOLOv8-n相比,mAP指标在TDD(traffic detection dataset)数据集上分别提升2.6%和0.2%,在Cmudsodd(coal mine underground drilling site object detection dataset)数据集上分别提升2.1%和4.4%。结论本文提出的长短程依赖特征金字塔网络解决了YOLOv7-tiny在特征融合阶段存在的相邻层特征融合时信息丢失问题和非相邻层特征信息差异问题,提升了YOLOv7-tiny方法的检测性能,并优于两种参数量相当的方法YOLOv5-s和YOLOv8-n。展开更多
文摘In this study,we propose Space-to-Depth and You Only Look Once Version 7(SPD-YOLOv7),an accurate and efficient method for detecting pests inmaize crops,addressing challenges such as small pest sizes,blurred images,low resolution,and significant species variation across different growth stages.To improve the model’s ability to generalize and its robustness,we incorporate target background analysis,data augmentation,and processing techniques like Gaussian noise and brightness adjustment.In target detection,increasing the depth of the neural network can lead to the loss of small target information.To overcome this,we introduce the Space-to-Depth Convolution(SPD-Conv)module into the SPD-YOLOv7 framework,replacing certain convolutional layers in the traditional system backbone and head network.This modification helps retain small target features and location information.Additionally,the Efficient Layer Aggregation Network-Wide(ELAN-W)module is combined with the Convolutional Block Attention Module(CBAM)attention mechanism to extract more efficient features.Experimental results show that the enhanced YOLOv7 model achieves an accuracy of 98.38%,with an average accuracy of 99.4%,outperforming the original YOLOv7 model.These improvements represent an increase of 2.46%in accuracy and 3.19%in average accuracy.The results indicate that the enhanced YOLOv7 model is more efficient and real-time,offering valuable insights for maize pest control.
文摘目的YOLOv7-tiny(you only look once version 7-tiny)成为实时目标检测领域的常用方法,由于其轻量化网络架构设计和较少的参数量,整个训练过程在单个网络中进行,检测速度快且不需要使用滑动窗口或候选区域,在资源受限、实时性要求高的任务中表现优异。然而,YOLOv7-tiny在特征融合阶段存在相邻层特征融合时信息丢失和非相邻层特征信息差异两个问题。为了解决上述问题,提出一种长短程依赖特征金字塔网络LSRD-FPN(long short range dependency feature pyramid network),并基于该网络对YOLOv7-tiny方法进行改进。方法LSRD-FPN包括两个关键组成部分:局部短程依赖机制SRD(short range dependency)和全局长程依赖机制LRD(long range dependency)。局部短程依赖机制通过改进上采样方式和引入注意力机制,有效缓解了特征融合过程中信息丢失的问题;全局长程依赖机制通过引入跨层连接模块,将主干网络的多尺度特征缩放、融合并分配到检测阶段的不同层级特征。LSRD-FPN不仅增强了模型的特征表达能力,而且提升了其在多尺度目标检测任务的性能表现。结果选用两个不同场景和规模的数据集进行实验。实验结果表明,相较于YOLOv7-tiny,本文方法的mAP分别取得1.3%和0.5%的性能提升。与参数量相当的YOLOv5-s和YOLOv8-n相比,mAP指标在TDD(traffic detection dataset)数据集上分别提升2.6%和0.2%,在Cmudsodd(coal mine underground drilling site object detection dataset)数据集上分别提升2.1%和4.4%。结论本文提出的长短程依赖特征金字塔网络解决了YOLOv7-tiny在特征融合阶段存在的相邻层特征融合时信息丢失问题和非相邻层特征信息差异问题,提升了YOLOv7-tiny方法的检测性能,并优于两种参数量相当的方法YOLOv5-s和YOLOv8-n。
