基于多尺度序列特征融合的林区害虫检测算法  

A Forest Pest Detection Algorithm Based on Multi-scale Sequence Feature Fusion

在线阅读下载全文

作  者:唐延龄 韩巧玲[1] 赵玥[1] 刘卫平[1] 郑一力[1] 赵燕东[1] 徐钐钐 Tang Yanling;Han Qiaoling;Zhao Yue;Liu Weiping;Zheng Yili;Zhao Yandong;Xu Shanshan(College of Technology,Beijing Forestry University Beijing 100091)

机构地区:[1]北京林业大学工学院,北京100091

出  处:《林业科学》2025年第2期21-30,共10页Scientia Silvae Sinicae

基  金:国家自然科学基金青年科学基金项目(32101590);国家自然科学基金面上项目(32071838);北京林业大学“5·5工程”科研创新团队项目(BLRC2023C05)。

摘  要:【目的】针对虫情测报灯采集到的林区害虫样本种类多、尺寸复杂及密集遮挡问题,提出了一种多尺度序列特征融合检测算法(MPD-YOLO),用于提升林区害虫检测的精确度,为林区害虫监测和防护提供了新的技术路径。【方法】首先,以采集于北京、河北、山西等地的林区害虫图像为基础,构建包含18类林区害虫的数据集。数据集涵盖不同场景下的害虫图像,为算法的训练与测试提供了可靠的数据支撑。其次,为提高小目标害虫的检测效果,利用3D卷积对小目标进行更为深入的尺度序列特征提取,MPD-YOLO方法引入尺度序列特征融合模块(SSFF),有效提升模型对小目标的感知能力。最后,为解决密集遮挡和多尺度害虫并存情况下的模型检测能力,MPD-YOLO方法提出了三重特征编码高效聚合模块(TFE-ELAN),通过将不同尺寸特征图进行特征融合,增强多层特征图之间信息的关联性,提升了模型的检测性能。【结果】在相同试验条件下,本研究在构建的数据集上对MPD-YOLO进行大量试验,并与主流的目标检测算法(YOLO系列、Faster-CNN等)进行对比分析。MPD-YOLO方法具有最佳的害虫识别精度,其F1-score和m AP分别达到88.43%和91.92%,比次优方法YOLOv8x分别高1.45%和1.22%。相比于原网络模型YOLOv7,MPD-YOLO的F1-score与mAP分别比YOLOv7高0.72%和2.8%,证明了本文所提方法在处理复杂目标检测任务中的优势。此外,本研究中消融试验,分析了尺度序列特征融合模块和三重特征编码高效聚合模块对整体性能的贡献,验证了各模块设计的有效性。【结论】本文提出的MPD-YOLO算法,有效提升了复杂环境下多尺度害虫的检测性能,在处理复杂环境下的检测任务时表现出较高的鲁棒性。【Objective】This study addresses the challenges of detecting forest pests,collected by pest monitoring lamps,that are characterized by high diversity,complex sizes,and dense occlusions.We propose a multi-scale pest detection algorithm(MPDYOLO)aimed at improving the accuracy of forest pest detection,thereby providing a new technological path for monitoring and protection of pests in forest areas.【Method】First,a dataset comprising images of forest pests from regions such as Beijing,Hebei,and Shanxi was constructed.This dataset includes 18 categories of forest pests from various scenarios,providing a robust foundation for algorithm training and evaluation.Second,to improve the detection of small pests,MPD-YOLO employs 3D convolutions to extract deeper scale-sequence features from small targets,and introduces a scale sequence feature fusion module to enhance the model’s sensitivity to small objects.Finally,to address the challenges posed by dense occlusions and the coexistence of multi-scale pests,the MPD-YOLO method incorporates a triple feature encoding effective long-range aggregation network.This module fuses features across different scales,enhancing the interaction between multi-layer feature maps and improving the model’s detection performance.【Result】Extensive experiments were conducted on the constructed dataset under identical test conditions,and the MPD-YOLO method was compared with mainstream object detection algorithms(YOLO series,Faster RCNN,etc.).MPD-YOLO demonstrated superior pest recognition accuracy,with F1-score and mAP of 88.43%and 91.92%,respectively,outperforming the second-best method,YOLOv8x,by 1.45%and 1.22%.Additionally,compared to the original YOLOv7 model,MPD-YOLO achieved improvements of 0.72%in F1-score and 2.8%in mAP,confirming its advantages in handling complex object detection tasks.Ablation studies further analyzed the contributions of the scale sequence feature fusion and triple feature encoding modules,validating the effectiveness of each design component.【Conclusi

关 键 词:林区害虫防控 林区生态系统 小目标检测 YOLOv7 特征融合 

分 类 号:S763.30[农业科学—森林保护学]

 

参考文献:

正在载入数据...

 

二级参考文献:

正在载入数据...

 

耦合文献:

正在载入数据...

 

引证文献:

正在载入数据...

 

二级引证文献:

正在载入数据...

 

同被引文献:

正在载入数据...

 

相关期刊文献:

正在载入数据...

相关的主题
相关的作者对象
相关的机构对象