摘要
针对当前飞机遥感图像目标检测算法的精度和实时性不能兼顾的问题,提出了基于SSD(Single Shot MultiBox Detector)的飞机遥感图像目标检测算法。首先使用经过改进后的深度残差网络替换SSD的骨架网络,对于特征图之间缺少特征信息关联和特征通道间缺少差异性权重值的问题,设计了一种含有特征感受野增强模块与注意力机制模块的新型特征金字塔网络。该网络用以融合不同层级的特征信息和训练特征通道间的权重系数,使得深层网络和浅层网络都得到结构层次丰富的融合特征,为后续网络的分类与定位提供了良好的前提。另外,在改进后的SSD算法中还使用了聚焦分类损失函数来解决正负样本不平衡的问题。在飞机遥感数据集上进行相关实验,精度均值达到92.45%,每秒帧率为35.6。结果表明,改进后的SSD算法能够同时兼顾高检测精度和实时性。
Aiming at the problem that the accuracy and real-time performance of current aircraft remote sensing image target detection algorithms can not be balanced,an aircraft remote sensing image target detection algorithm based on single shot multibox detector(SSD)is proposed.Firstly,the improved deep residual network is used to replace the skeleton network of SSD.For the problems of lack of feature information association between feature maps,and different channels from feature maps have no weight values,this paper designs a new feature pyramid network with feature receptive field enhancement module and attention mechanism module.The network is used to fuse the feature information of different levels and the weight coefficients between the training feature channels,so that both the deep network and the shallow network can obtain fusion features with rich structural levels,which provides a good prerequisite for the classification and positioning of the subsequent network.In addition,the focus classification loss function is also used in the improved SSD algorithm to solve the problem of imbalance between positive and negative samples.Related experiments are carried out on the aircraft remote sensing data set,and the average accuracy reaches 92.45%,and the frame per second is 35.6.The results show that the improved SSD algorithm can balance high detection accuracy and real-time performance at the same time.
作者
王浩桐
郭中华
WANG Hao-tong;GUO Zhong-hua(School of Physics and Electronic and Electrical Engineering, Ningxia University, Yinchuan 750021, China;Ningxia Key Laboratory of Desert Information IntelligentPerception, Yinchuan 750021, China)
出处
《液晶与显示》
CAS
CSCD
北大核心
2022年第1期116-127,共12页
Chinese Journal of Liquid Crystals and Displays
基金
宁夏自然科学基金(No.2020AAC03026)
宁夏大学研究生创新研究项目(No.GIP2020075)。
关键词
目标检测
遥感图像
特征融合
感受野增强
注意力机制
target detection
remote sensing image
feature fusion
attention mechanism
receptive field enhancement
作者简介
王浩桐(1995-),男,宁夏青铜峡人,硕士研究生,2018年于重庆邮电大学获得学士学位,主要从事计算机视觉方面的研究,E-mail:562287297@qq.com;通信联系人:郭中华(1973-),男,宁夏银川人,博士,教授,2008年于西北工业大学获得博士学位,主要从事图像处理、机器视觉方面的研究,E-mail:guozhh@nxu.edu.cnmail:guozhh@nxu.edu.cn。
