改进特征金字塔网络的小目标检测

Improved Feature Pyramid Network for Small Object Detection

由于小目标可视化特征少、难以定位,故许多小目标检测方法基于特征金字塔网络(FPN)进行多尺度融合丰富各特征层的信息。然而,FPN只关注特征局部相关性并且使用逐元素相加操作融合不同的特征层忽略了不同特征层感受野的不同。因此,提出增强上下文特征金字塔网络(ECFPN),设计了上下文信息增强(CIE)模块增强上下文信息,注意力引导特征融合(AGFF)模块融合高层特征图和低层特征图。实验结果表明,ECFPN在VOC2012数据集上的AP 0.5、AP S分别达到75.05%和19.48%,在NWPU VHR-10数据集上的AP 0.5、AP S分别达到93.48%和45%,具有良好的小目标检测性能。

Because there are few visual features of small targets and it is difficult to locate them,many small object detection methods are based on Feature Pyramid Network(FPN)to perform multi-scale fusion to enrich the information of each feature layer.However,FPN only focuses on the local correlation of features and uses element-wise addition operations to fuse different feature layers,ignoring the differences in the receptive fields of different feature layers.Therefore,the Enhance Context Feature Pyramid Network(ECFPN)is proposed,the Context Information Enhancement(CIE)is designed to enhance context information,and the Attention Guided Feature Fusion(AGFF)fuses high-level feature maps and low-level feature maps.The experimental results show that the ECFPN has AP 0.5 and AP S of 75.05%and 19.48%respectively on VOC2012 dataset,and of 93.48%and 45%respectively on NWPU VHR-10 dataset,which has good small target detection performance.