摘要
针对密集目标检测任务中由于多尺度和遮挡现象导致漏检的问题,文中提出一种改进YOLOv8n的方法(DERFYOLO)。对于主干网络,首先使用C2f-RFCBAMConv模块替换卷积和C2f模块,利用感受野注意力机制优化的空间特征并结合通道注意力机制以提高改进算法特征提取的能力;其次设计EMBSFPN模块对颈部进行优化,实现小目标信息的跨尺度连接和特征融合,通过更换上采样模块使得算法能够在保证一定效果的同时保持高效性;最后使用DyHead代替原检测头进而引入注意力机制以提升小目标检测的精度。在VisDrone2019数据集上的实验结果表明,DERF-YOLO的mAP@0.5、mAP@0.5:0.95分别达到了30.9%和17.7%,相比于YOLOv8n算法分别提高了4.0%和2.7%,参数量和浮点运算量分别为2.94×10~6和9.6 GFLOPs,参数量相比于原始算法降低了2%,运算量仅仅增加了18%。该算法的精度高于其他同类算法,且满足监测需求,可以有效地应用于无人机航拍平台上的目标检测任务。
In view of the missed detection caused by multi-scale and occlusion in dense object detection tasks,this paper proposes an algorithm on the basis of the improved YOLOv8n,and the algorithm is named as DERF-YOLO.In the backbone network,the C2f-RFCBAMConv module is used to replace convolution and C2f modules,and the receptive field attention mechanism is utilized to optimize spatial features and then is combined with channel attention mechanisms to enhance feature extraction capabilities of the improved algorithm.The EMBSFPN module is designed to optimize the neck part,achieving cross-scale connection and feature fusion of small object information.By replacing the up-sampling module,the algorithm can maintain efficiency while ensuring certain effects.Finally,DyHead is employed instead of the original detection head,and the attention mechanisms are introduced to improve the accuracy of small object detection.Experimental results on the VisDrone2019 dataset show that the mAP@0.5 and mAP@0.5:0.95 of the DERF-YOLO reach 30.9%and 17.7%,respectively,representing improvements of 4.0%and 2.7%,respectively,in comparison with the YOLOv8n algorithm.The parameter count and floating-point operations of the DERF-YOLO are 2.94×106 and 9.6 GFLOPs,respectively,with a 2%reduction in parameters and only an 18%increase in computation burden in comparison with those of the original algorithm.Its accuracy is higher than that of the other similar algorithms and it meets the monitoring requirements and can be effectively applied to the object detection tasks on UAV aerial photography platform.
作者
曲文龙
陈勇
QU Wenlong;CHEN Yong(School of Mathematics and Computer Science,Shaanxi University of Technology,Hanzhong 723001,China)
出处
《现代电子技术》
北大核心
2026年第1期77-85,共9页
Modern Electronics Technique
基金
陕西省自然科学基础研究计划项目(2024JC-YBQN-0725)。
