基于多层级自注意力增强的遥感目标检测

Remote Sensing Target Detection Based on Multilevel Self-Attention Enhancement

随着遥感图像分辨率的不断提高,遥感图像目标检测技术获得了更广泛的关注。针对遥感图像中背景复杂噪声多、目标方向任意且目标尺寸变化大等问题,提出一种基于多层级局部自注意力增强的遥感目标检测算法。首先,在Oriented R-CNN骨干网络中引入Swin Transformer特征提取模块,使用具有移位窗口操作和层次设计的Transformer模块对特征提取的语义信息进行多层级局部信息建模。其次,使用Oriented RPN生成高质量的有向候选框。最后,将高斯分布之间的Kullback-Leibler divergence(KLD)作为回归损失函数,使得参数梯度能够根据对象的特征得到动态调整,更加准确地进行检测框的回归。所提算法在DOTA数据集和HRSC2016数据集上的平均精度均值(mAP)分别达77.2%和90.6%,和Oriented R-CNN算法相比,mAP分别提高了1.8个百分点和0.5个百分点。实验结果表明,所提算法能够有效地提高遥感图像目标检测精度。

Remote sensing image target detection technology has gained considerable attention with the improvement of remote sensing image resolution.This thesis proposes a remote sensing target detection algorithm based on multilevel local self-attention enhancement to solve such problems as complex background noise,arbitrary target direction,and large changes in target size in remote sensing images.First,the proposed algorithm adopts the Swin Transformer feature extraction module in an Oriented region-based convolutional neural network(R-CNN) backbone network,and the multilevel local information of feature-extracted semantic information is modeled using the Transformer module with shifted window operations and hierarchical design.Second,Oriented RPN is used to generate high-quality directed candidate boxes.Finally,the Kullback-Leibler divergence(KLD) between Gaussian distributions is regarded as the regression loss function,allowing the parameter gradient to be dynamically adjusted based on the object's characteristics for more accurate regression of the detection boxes.The mean average precision(mAP) of the proposed algorithm reaches 77.2% and 90.6% on the DOTA dataset and HRSC2016 dataset,respectively,and it is increased by 1.8 percentage points and 0.5 percentage points compared with the Oriented R-CNN algorithm.The results reveal that the proposed algorithm can effectively advance the target detection accuracy of remote sensing images.