针对小尺寸列车风管与折角塞门的检测问题,提出一种多尺度特征增强的YOLOv5s检测算法.首先在主干网络Backbone增加了一个大尺寸特征图输出分支,使用C2f(CSP bottleneck with 2 convolution and forward)模块丰富新增特征图的梯度流信息;然后在Neck部分使用自适应空间特征融合增强特征金字塔模型FPN(Feature Pyramid Net-work)的多尺度特征融合能力,并使用路径聚合模型PAN(Path Aggregation Network)提高目标定位能力;检测头Head拥有 4 种尺度的特征图,其中新增特征图的尺寸最大,用于增强小尺寸目标的检测能力.实验结果表明,多尺度特征增强的YOLOv5s比原来的YOLOv5s在平均精度mAP(mean Average Precision)上提升了 1.5%,优于其他检测器,提升了小尺寸风管和折角塞门的检测能力.
Detection of Duct Picking and Angle Cock of Train Based on Multi-scale Feature
To improve the performance for detecting the duct pickings and angle cocks of train,an improved YOLOv5s based on multiple scale feature is proposed in this paper.Firstly,a new feature output is added at the backbone of YOLOv5s,where C2f block is exploited to enrich the gradient information and to detect the small object.The size of newly added feature map is doubled compared with that of original map.Then,an enhanced method for feature fusions is pro-posed at the Neck.Adaptively spatial feature fusion(ASFF)was utilized to improve the effectiveness of fused multiple scales of feature.Finally,the enhanced FAN with PAN framework was incorporated to fuse the feature maps of different size.Experimental results show that the proposed YOLOv5s achieves about 1.5%higher of mAP than the initial YOLOv5s,outperforming other compared methods.
object detectionduct picking of trainangle cockmulti-scale featureadaptively spatial feature fusion
NETL
NSTL
万方数据
