结合非对称卷积组与通道注意力的遥感影像建筑检测方法

Building Detection in Remote Sensing Images Combining Asymmetric Convolution Groups and Channel Attention

针对以遥感卫星影像为基础的建筑物检测问题,提出一种基于单阶段回归检测器的建筑物检测模型。在模型特征提取网络中通过非对称卷积核组进行特征提取操作,同时采用通道注意力层进一步筛选建筑物目标的细节特征;在特征增强网络中,使用特征金字塔与PAN结构层组合对来自提取网络中的特征图进行融合增强,最终生成4个不同尺度的特征图送入检测端。实验结果表明,本文模型在精度均值方面优于几组主流深度学习检测模型,能够对卫星遥感影像中多尺寸、多角度的建筑物目标精准检测,对不同环境下目标也具有很好的鲁棒性,在实验环境下能够达到实时级别的检测速度。

Aiming at the problem of building detection in remote sensing satellite images,this paper proposes a building detection model based on a single-stage regression detector.In the model feature extraction network,the feature extraction operation is performed through asymmetric convolution kernel groups,and the channel attention layer is used to further filter the detailed features of the building targets;in the feature enhancement network,the combination of the feature pyramid and the PAN structure layer is used for fusion enhancement of the feature maps in the network,and finally four feature maps of different scales are generated and sent to the detection end.The experimental results show that the model in this paper is superior to several mainstream deep learning detection models in terms of average accuracy,can accurately detect building targets of multiple sizes and angles in satellite remote sensing images,has good robustness to targets in different environments,and has the level of real-time detection in the experimental environment.