基于改进特征金字塔的目标检测方法

Object Detection Method Based on Improved Feature Pyramid Networks

为了缓解多尺度目标特征信息不足的问题,受BiFPN(Bi-directional Feature Pyramid Network)网络的启发,在网络模型的Neck部分提出了一种反转N型特征金字塔结构即IN-FPN(Invert N-Feature Pyramid Network),对网络的多尺度特征融合结构加以优化,其带有侧向连接的层次结构,将特征经过2次自上而下和1次自下而上的双向融合,使得物体的浅层和深层特征充分融合,相互促进.同时,考虑不同尺度特征融合时的贡献不同,给每一个尺度添加可自适应学习权重ωi.此外,为了解决网络退化问题,进一步提升网络性能,参考残差网络结构,增加了含有Block模块的路径.实验结果表明,所提方法在COCO 2017数据集和VisDrone 2019数据集上其平均精度(AP)值分别达到了53.02%和25.21%,比基准模型均有所提升,验证了该方法的有效性.

To solve the problem of insufficient feature information for multi-scale targets,inspired by the BiFPN network,we proposed IN-FPN,an inverted N-type FPN structure in the Neck section of the network model.In addition,the multi-scale feature fusion structure of the network was optimized,with a horizontally connected hierarchical structure.The features are fused twice from top to bottom and once from bottom to top in a bidirectional manner,allowing shallow and deep features of the object to fully fuse and promote each other.At the same time,for the different functions of feature fusion at different scales,adaptive learning weights are added to each scaleωi.Moreover,in order to solve the problem of network degradation and improve network performance,a path containing a Block module was added based on the residual network structure.The experimental results showed:With this method,the AP values reached 53.02%and 25.21%on the COCO 2017 dataset and VisDrone 2019 dataset.They were both improved,compared with the benchmark model,verifying the effectiveness of the method.