一种用于目标检测的位置加权特征金字塔网络

A Position-Weighted Feature Pyramid Network for Object Detection

针对传统特征金字塔网络的融合方式会导致混叠效应、高级语义信息被稀释和低级空间结构信息局限访问的问题,提出一种新的位置加权特征金字塔网络。首先设计了一个位置加权模块,从不同等级的特征图的所有通道中学习得到位置权重并赋给各等级特征图实现位置加权,进而引导融合过程;然后引入了一个金字塔池化模块,提取包含特征图不同大小区域的上下文信息的多尺度特征,并与位置加权模块相结合。实验表明,提出的位置加权特征金字塔网络相比于传统特征金字塔网络对不同类型的目标检测网络的检测精度皆有提升效果。

For the problem which is aliasing effects,dilution of high-level semantic information and limited access to low-level spatial structure information caused by traditional feature pyramid network fusion.A new position-weigh-ted feature pyramid network is proposed.Firstly,a position-weighted module was designed to get position weight from all channels of different levels of feature maps.The position weight was assigned to the feature map of each level,ac-cordingly,each position was weighted and then served as a guidance to boost the fusion process.Next,a pyramid poo-ling module was introduced to extract multi-scale features containing contextual information of different size regions of the feature map and combined with the position-weighted module.Experiments show that compared with the tradition-al feature pyramid network,the position-weighted feature pyramid network can improve the detection accuracy of dif-ferent types of detection networks.