摘要
针对现有方法在深层网络中未能充分利用浅层特征,从而导致单应性估计性能受限的问题,提出了一种基于Unet的红外与可见光图像单应性估计方法。分别利用特征提取器和掩码预测器提取图像对的特征图和掩码。再分别把图像对的特征图和掩码进行相乘,以获取加权特征图,从而更好地突出特征图中对单应性估计有很大贡献的内容。把图像对的加权特征图进行通道级联以作为单应性估计器的输入,从而得到单应性矩阵。该单应性估计器的主干是Unet,它能有效地在深层网络中整合浅层低级特征。大量实验结果表明,Unet的引入可有效提升单应性估计性能,角点误差从5.25显著下降至5.17。
To address the issue that existing methods fail to effectively utilize shallow features in deep networks,which results in the limited performance of single strain estimation,a homography estimation method for infrared and visible images based on Unet is proposed.First,the feature maps and masks of the image pairs are extracted using a feature extractor and a mask predictor,respectively.Second,the feature maps and masks of the image pairs are multiplied respectively to obtain the weighted feature maps,to highlight the contents of the feature maps that contribute significantly to the homography estimation.Finally,the weighted feature maps of the image pairs are channel cascaded to serve as the input for the homography estimator to obtain the homography matrix.The backbone of this homography estimator is Unet,which can effectively integrate shallow low level features in deep networks.Extensive experimental results show that introduced Unet can effectively improve the homography estimation performance,and the corner error significantly decreased from 5.25 to 5.17.
作者
王星怡
罗银辉
吴岳洲
魏嗣杰
WANG Xing-yi;LUO Yin-hui;WU Yue-zhou;WEI Si-jie(Civil Aviation Flight University of China,Guanghan 618000,China)
出处
《航空计算技术》
2023年第4期51-55,共5页
Aeronautical Computing Technique
基金
国家重点研发计划项目资助(2021YFF0603904)
四川省科技计划项目资助(2022YFG0027)
中央高校基本科研业务费基金项目资助(ZJ2022-004
ZHMH2022-006)
中国民用航空飞行学院研究生科研创新基金项目资助(X2023-27)。
关键词
单应性估计
红外图像
可见光图像
homography estimation
infrared image
visible image
