结合卷积核组与位置注意力的遥感影像建筑物检测方法

The detection method for remote sensing image buildings combined convolution kernel group with position attention

针对高分辨率遥感影像中建筑物样本尺寸较小且分布不均,使用现有方法难以实现对建筑物目标高精度检测的问题,提出一种基于卷积核组与位置注意力的单阶段建筑物检测模型。在模型骨干网络中,使用并联卷积核组进行特征提取,同时引入位置卷积模块为正样本特征赋权来提高模型学习效率;为强化特征图内的信息丰富程度,使用带有密集连接的特征增强网络对骨干网络输出的特征图进行强化,最终将4个不同尺度特征图送入检测端。试验结果表明:提出的模型在检测精度方面优于参照模型,并且在不同场景下表现出较好的泛化能力和小目标检出能力,同时能够在测试环境下实现对建筑物目标的实时检测。

In view of such problems as the small size,uneven distribution of building samples in high-resolution remote sensing images and using existing methods is difficult to achieve high-precision detection of building targets,a single-stage building detection model based on convolution kernel group and positional attention is proposed in this study.In the model backbone network,the parallel convolution kernel group is used for feature extraction,and the position convolution module is introduced to weight the positive sample features to improve the model learning efficiency.In order to enhance the information richness in the feature map,the feature enhancement network with dense connections is used to strengthen the feature map output from the backbone network,and finally four different scale feature maps are sent to the detection end.The experimental results show that the proposed model is superior to the reference model in terms of detection accuracy and has good generalization ability and small target detection ability in different scenarios.At the same time,it can achieve real-time detection for building targets in the test environment.