基于上下文增强和特征提纯的小目标检测网络  

Small Target Detection Network Based on Context Augmentation and FeatureRefinement

作　　者：肖进胜[1] 赵陶 周剑[2] 乐秋平 杨力衡 Xiao Jinsheng;Zhao Tao;Zhou Jian;Le Qiuping;Yang Liheng(School of Electronic Information,Wuhan University,Wuhan 430072;State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing(Wuhan University),Wuhan 430079)

机构地区：[1]武汉大学电子信息学院,武汉430072 [2]测绘遥感信息工程国家重点实验室(武汉大学),武汉430079

出　　处：《计算机研究与发展》2023年第2期465-474,共10页Journal of Computer Research and Development

基　　金：国家自然科学基金青年科学基金项目(42101448);中国科学院光电信息处理重点实验室开放课题基金项目(OEIP-O-202009)。

摘　　要：微小目标的纹理模糊、包含特征少,是目标检测领域的难点.针对小目标检测提出一种新的上下文增强模块(context augmentation module,CAM)和特征提纯模块(feature refinement module,FRM)相结合的特征金字塔复合结构.利用多尺度空洞卷积的特征融合,补充网络中的上下文信息;引入通道和空间的特征提纯机制来抑制多尺度特征融合后的冲突信息,防止小目标淹没在冲突信息中;同时,引入复制—缩小—粘贴(copy-reduce-paste)的数据增强方法提高小目标的占比,使训练时小目标对损失值的贡献更大,训练更加平衡.由实验结果可知,所提出的算法在VOC数据集上目标检测的平均精度均值(Mean Average Precision,mAP)达到了83.6%(交并比为0.5);对小目标检测的AP值达到了16.9%(交并比为0.5~0.95),比YOLOV4,CenterNet,RefineDet的分别提高3.9%,7.7%和5.3%.在TinyPerson数据集上小目标检测的AP值为55.1%,比YOLOV5,DSFD的分别提高0.8%和3.5%.Small objects contain few and fuzzy features,which is a hard problem in the field of object detection.The poor performance of small object detection is mainly caused by the limitation of the network and the imbalance of the training dataset.A novel feature pyramid composite structure constructed by context augmentation module(CAM)and feature refinement module(FRM)is proposed.The feature fusion of multi-scale dilated convolution is applied to generate features on different receptive fields,and then the features are added to detection network to supplement context information.The channel and space feature refinement mechanism is introduced to suppress the conflict information generated by multi-scale feature fusion and prevent small objects from being submerged in the conflict information.Besides,a data augmentation method called copy-reduce-paste is proposed to increase the proportion of small targets,so that the contribution of small targets to the loss value during training is greater and the training is more balanced.Experimental results show that the Mean Average Precision(mAP)of object detection on the VOC dataset of the proposed network is 83.6%(IOU is 0.5).The AP value of small target detection is 16.9%(IOU changes from 0.5 to 0.95),which is 3.9%,7.7%and 5.3%higher than that of YOLOV4,CenterNet and RefineDet,respectively.The AP value of small target detection on TinyPerson dataset is 55.1%,which is 0.8%and 3.5%higher than that of YOLOV5 and DSFD,respectively.

关 键 词：小目标检测 上下文增强 特征提纯 空洞卷积 数据增强 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]