一种特征由深到浅叠加融合的多尺度目标检测模型  

A Multi-Scale Target Detection Method Based on Superposition and Fusion of Features from Deep to Shallow

作　　者：陈海燕[1] 李春尧 CHEN Haiyan;LI Chunyao(College of Computer and Communication,Lanzhou University of Technology,Lanzhou Gansu 730050,China)

机构地区：[1]兰州理工大学计算机与通信学院,甘肃兰州730050

出　　处：《传感技术学报》2022年第10期1375-1381,共7页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金项目(6216010086)。

摘　　要：基于深度学习的特征金字塔网络(Feature Pyramid Networks,FPN)仅采用一次上采样与相邻层特征融合的方法,存在浅层网络与深层网络特征关联性不强,多层网络特征融合不充分的问题,影响多尺度目标检测精度。对此,将主干网络中提取的特征进行由深到浅的叠加融合,并对特征金字塔中得到的特征进行补充叠加融合。此外,为进一步提高检测器对目标特征的识别能力,对每次叠加融合后得到的特征通过non-local网络进行特征增强。以PASCAL VOC为数据集的目标检测实验结果表明,所提目标检测模型对数据集中所有类别目标的平均检测精度(mean Average Precision,mAP)为80.6%,对行人类别的检测精度(Average Precision,AP)为81.3%,较FPN网络分别提高了2.4%和2.8%,有效提高了多尺度目标检测精度。Feature Pyramid Networks(FPN)based on deep learning only uses on time of upsampling and fusion of adjacent layer features,and it has the problems of weak correlation between shallow network and deep network features and insufficient fusion of multi-layer network features,which affect the accuracy of multi-scale target detection.In order to solve these problems,the features extracted from the backbone network are superimposed and fused from deep to shallow,and then the features obtained in the feature pyramid are superimposed and fused for the second time.In addition,in order to further improve the detector’s ability to recognize the target features,the non-local network is used to enhance the features obtained after each stacking and fusion.The experimental results show that the average detection accuracy of the target detection model is 80.6%in all categories of PASCAL VOC,and the detection accuracy in the pedestrian category of PASCAL VOC is 81.3%,which are 2.4%and 2.8%higher than those of the FPN network respectively.Therefore,the model proposed effectively improves the multi-scale target detection accuracy.

关 键 词：深度学习 目标检测 特征融合 多尺度 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]