联合多尺度与注意力机制的遥感图像目标检测

遥感图像存在背景复杂、目标尺度差异大且密集分布等不足，为提高现有算法的检测效果提出联合多尺度与注意力机制的遥感图像目标检测算法. 改进空洞空间金字塔池化模块，增大不同尺寸图像的感受野；提出注意力模块用于学习特征图通道信息和空间位置信息，提升算法对复杂背景下遥感图像目标区域的特征提取能力；引入加权双向特征金字塔网络结构与主干网结合来增进多层次特征的融合；使用基于距离的非极大值抑制方法进行后处理，改善检测框易重叠的问题. 在DIOR和NWPUVHR-10数据集上的实验结果表明：所提算法的平均精度均值mAP分别达到71.6%和91.6%，相比于主流的YOLOv5s算法分别提升了2.9%和1.5%. 所提算法对复杂遥感图像取得了更好的检测效果.

Remote sensing image target detection combining multi-scale and attention mechanism

Remote sensing images have deficiencies such as complex backgrounds, significant differences in target scales, and dense distribution, resulting in poor detection of existing algorithms. A remote sensing image object detection algorithm that combined multi-scale and attention mechanisms was proposed. The receptive field of images of different sizes improved the atrous spatial pyramid pooling module. An attention module was proposed to improve the feature extraction ability for target regions of remote sensing images under complex backgrounds by learning the feature map channel information and the spatial location information. A weighted bidirectional feature pyramid network structure was introduced to combine with the backbone network to improve the fusion of multi-level features. A distance-based non-maximum suppression method was used for postprocessing, which improved the problem of easy overlapping of detection frames. Experimental results on DIOR and NWPU VHR-10 datasets showed that the mean average precision (mAP) of the proposed algorithm reached 71.6% and 91.6%, which were 2.9% and 1.5% higher than those of the mainstream YOLOv5s algorithm respectively. The algorithm achieved good detection results for complex remote sensing images.