摘要
目标检测是遥感影像处理中一项基础性和常规性的工作。本文基于YOLOX(you only look once X)网络进行改进,设计了一种针对遥感影像目标的检测算法。首先在特征提取模块PANet(path aggregation network)中加入自适应空间特征融合(adaptively spatial feature fusion, ASFF)网络,针对目标检测中尺度不一致的细部特征进行深入挖掘。其次,设计了基于ECA(efficient channel attention)的特征提取模块,高效通道交互在更加关注特征图中正样本特征信息的同时降低了模型的复杂性。再次,为了避免过拟合造成梯度消失、激活效果弱的问题,提出使用swish激活函数。最后,在DOTA(dataset for object detection in aerial images)上进行实验,通过消融实验定性分析、通过对比实验定量验证了本文算法的最佳机制和有效性。结果表明:在添加ASFF和ECA机制并且优化swish激活函数的前提下,改进网络模型的全类平均正确率(mean average precision, mAP)达74.42%,较原始网络提升了12.75%;与当前应用广泛的目标检测算法Mobilenet-YOLOv4、YOLOv4、YOLOv5、YOLOX相比,本文算法实现了mAP 11.42%~17.84%的提升。
Target detection is a fundamental and routine task in remote sensing image processing.In this paper,we design a target detection algorithm for remote sensing images based on the YOLOX network.Firstly,ASFF is added to the feature extraction module PANet to deeply mine the fine features with inconsistent scale in target detection.Secondly,an ECA-based feature extraction module is designed with efficient channel interaction to reduce the complexity of the model while paying more attention to the positive sample feature information in the feature map.Then,to avoid the problem of gradient disappearance and weak activation effect caused by overfitting,the use of swish activation function is proposed.Finally,experiments are conducted on DOTA to verify the best mechanism and effectiveness of the improved method through qualitative analyses of the ablation experiments and quantitative comparison experiments.With the addition of ASFF and ECA mechanisms and the optimization of the swish activation function,the improved network model achieves an mAP of 74.42%,an improvement of 12.75%over the original network.Compared with the current widely used target detection algorithms Mobilenet-YOLOv4,YOLOv4,YOLOv5 and YOLOX,the proposed algorithm achieves an improvement of 11.42%-17.84% in mAP accuracy.
作者
李美霖
芮杰
金飞
刘智
林雨准
Li Meilin;Rui Jie;Jin Fei;Liu Zhi;Lin Yuzhun(Institute of Geospatial Information,PLA Strategic Support Force Information Engineering University,Zhengzhou 450001,China;61206 Troops,Beijing 100042,China)
出处
《吉林大学学报（地球科学版）》
CAS
CSCD
北大核心
2023年第4期1313-1322,共10页
Journal of Jilin University：Earth Science Edition
基金
国家自然科学基金项目(41601507)。
