摘要
基于端到端的检测框架,使用多空洞率卷积核组作为特征提取模块,并在不同特征提取层间设置了密集连接,来加强不同尺度特征图内的信息复杂度;以多尺度特征图融合为基础,构建了4个输出层的特征图上采样金字塔,最后通过数据增强提高了训练集内目标的表达能力。测试结果表明,本文方法在测试数据集上能够达到较高的检测精度,体现了良好的实时检测能力,并且对不同背景下多角度的房屋目标具有很好的泛化性能。该方法在城市违章建筑监管与智慧城市建设等领域具有较高的实用价值。
Based on an end-to-end detection framework,this paper uses a multi-dilation rate convolution kernel group as a feature extraction module,and sets up dense connections between different feature extraction layers to enhance the information complexity in feature maps of different scales;Based on multi-scale feature map fusion,the feature map upsampling pyramid of four output layers is constructed,and finally the expression ability of the target in the training set is improved through data augmentation.The test results show that the method in this paper can achieve high detection accuracy and real-time detection ability on the test data set,and has good generalization performance for multi-angle house targets in different backgrounds.This method has high practical value in the fields of urban illegal building supervision and smart city construction.
作者
刘瑶
亢玮
赵占营
LIU Yao;KANG Wei;ZHAO Zhanying(Anhui Geohold Technology Co.,Ltd.,Beijing 100020,China;Beijing Tianxia Map Data Technology Co.,Ltd.,Beijing 100011,China;Piesat Information Technology Co.,Ltd.,Beijing 100195,China)
出处
《测绘与空间地理信息》
2024年第4期149-152,共4页
Geomatics & Spatial Information Technology
关键词
遥感影像
建筑物检测
空洞卷积核
密集连接
多尺度特征金字塔
remote sensing image
building detection
hollow convolution kernel
dense connection
multi-scale feature pyramid
