摘要
针对基于YOLOv5算法的小目标的检测精度低、密集目标识别效果差的问题,提出了一种改进YOLOv5目标检测框架。YOLOv5的主干网络加入卷积块注意力模块(Convolutional block attention module,CBAM)可提高特征提取能力,增强网络对图像纹理的感知能力,使小目标获取更多关注。为了解决密集目标检测的漏检问题,YOLOv5的颈部网络使用加权双向特征网络(Bidirectional feature pyramid network,BiFPN)替代原有的像素聚合网络(Path aggregation network,PAN),通过权值共享的方式实现多尺度特征融合。采用EIoU作为模型的边界框回归损失函数,提高了边界框回归性能,加快网络收敛速度。在DOTA数据集上,实验验证了YOLOv5的改进结果,此方法的mAP为80.0%,能够检测更多的目标,相较于YOLOv5,改进YOLOv5的mAP提升了5.2%。
For the problem of low detection accuracy of small targets and poor recognition of dense targets based on YOLOv5 algorithm,the improved YOLOv5 target detection framework is proposed.The backbone network of YOLOv5 adds the Convolutional block attention module(CBAM)to enhance the network's ability to perceive the image texture so that the small targets get more attention.To solve the leakage problem of dense target detection,the neck network of YOLOv5 uses Bidirectional feature pyramid network(BiFPN)instead of the Path aggregation network(PAN)to realize multi-scale feature fusion by weight sharing.EIoU is used as the bounding box regression loss function of the model to strengthen the performance of bounding box regression and accelerate the network convergence.Experimental results on the DOTA dataset validate the improvement of YOLOv5.The mAP of the enhanced method is 80.0%,and the improved algorithm is able to detect more targets.Compared to YOLOv5,the mAP of the enhanced YOLOv5 is improved by 5.2%.
作者
刘国新
朱福珍
巫红
LIU Guoxin;ZHU Fuzhen;WU Hong(College of Electronic Engineering,Heilongjiang University,Harbin 150080,China)
出处
《黑龙江大学自然科学学报》
CAS
2024年第1期109-115,共7页
Journal of Natural Science of Heilongjiang University
基金
黑龙江省省属高校基本科研业务费项目(2023-KYYWF-1436,2022-KYYWF-1090)
国家自然科学基金资助项目(61601174,62341503)
黑龙江省“双一流”学科协同创新成果孵化项目(LJGXCG2023-046)
黑龙江大学横向课题项目(2023230101001032)。
