摘要
为提高卷积神经网络目标检测模型精度并增强检测器对小目标的检测能力,提出一种脱离预训练的多尺度目标检测网络模型。采用脱离预训练检测网络使其达到甚至超过预训练模型的精度,针对小目标特点设计新的Deformable-ScratchNet网络模型,调整网络结构并融合浅层信息以提高对小目标的检测性能。实验结果表明,与Faster-RCNN等经典网络模型相比,该模型在PASCAL VOC数据集和自制遥感军事目标数据集上的检测精度更高。
In order to improve the accuracy of the target detection model using convolutional neural network and enhance the detection ability of the detector for small targets,this paper proposes a multi-scale target detection network model trained from scratch.The detection network is trained from scratch to increase its accuracy to the level of pre-trained models or even higher.Then a new Deformable-ScratchNet network model is designed according to the characteristics of small targets.Its network structure is adjusted,and shallow information is integrated with the model to improve the detection performance of small targets.Experimental results show that compared with Faster-RCNN and other classic network models,the proposed model has higher detection accuracy on the PASCAL VOC data set and self-made remote sensing image of military target data set.
作者
包壮壮
赵学军
王明芳
董玉浩
庞梦洋
黄林
贺刚
BAO Zhuangzhuang;ZHAO Xuejun;WANG Mingfang;DONG Yuhao;PANG Mengyang;HUANG Lin;HE Gang(Department of Basic Science,Air Force Engineering University,Xi’an 710051,China;Unit 93861 of Chinese People’s Liberation Army,Xiangyang,Shaanxi 713800,China;Unit 32055 of Chinese People’s Liberation Army,Nanjing 210046,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2020年第6期248-255,共8页
Computer Engineering
基金
国家自然科学基金(61472443)。
关键词
脱离预训练
可变卷积
小目标检测
多尺度目标
遥感图像
trained from scratch
variable convolution
small target detection
multi-scale target
remote sensing image
