基于自适应校准和多分支注意力的遥感目标检测

Remote sensing object detection based on adaptive calibration and multi-branch attention

面向复杂多变的遥感场景下目标检测易受干扰的问题,提出了结合自校准模块和D_Triplet Attention的任意方向目标检测模型SD-Centernet。该方法在网络结构中引入旋转角度,为检测框提供角度信息。在Dlanet特征提取网络中引入self-Calibrated模块,通过自适应校准操作融合来自两个不同空间尺度的信息,增大输出特征的感受野。同时为了加强图像局部信息的聚焦,引入D_Triplet Attention,更好的解决了跨维度交互问题。SD-Centernet在HRSC-2016数据集上的检测精度达到86.25%,检测速度达到14.9帧/秒,有效提高了遥感航拍中多方位目标的检测效果。

In order to overcome the problem of vulnerable disturbance of target detection in complex and variable remote sensing situations,this paper suggests SD-Centernet,which is an optional orientation target detection model combining self-Calibrated module and D_Triplet Attention.This new method introduces rotation angle in the network structure,which provides angular information to the detection box.Self-Calibrated module is introduced into the Dlanet feature extraction network to increase the perceptual field of the output features by fusing information from two different spatial scales through an adaptive calibration operation.Meanwhile,D_Triplet Attention is introduced to enhance the focus of image-based local information,which better solves the cross-dimensional interaction problem.86.25%detection accuracy and 14.9 fps detection speed have been achieved on SD-Centernet in HRSC-2016 Dataset,that effectively improves the multi-directional target detection in remote sensing aerial photography.