遥感图像密集小目标全方位精准检测算法

Omnidirectional accurate detection algorithm for dense small objects in remote sensing images

针对遥感图像中目标排列密集且方向不相同,导致现有检测算法难以准确定位实例目标的问题,提出了一种遥感图像密集小目标全方位精准检测算法。首先,为提升特征提取能力,在主干网络的残差结构中引入Meta-ACON激活函数,自适应地学习信道特征的重要性;其次,提出一种加强连接特征金字塔网络,重新设计了用于深浅层特征融合的侧向连接部分,并在同层次特征图输入与输出之间添加了跳跃连接,丰富特征语义信息;再次,引入角度预测分支,使用环形平滑标签方法将角度回归问题转化为分类问题,在实现目标框旋转的同时解决了旋转框边界突变的问题;最后,设计针对旋转检测框的后处理方法(Rotate-Soft-NMS),通过抑制检测框的置信度去除相邻的重复旋转检测框。在DOTA数据集上的实验结果表明:该算法的平均精度均值达到76.15%,相比于基准模型YOLOv5m提升了5.22%,与其他先进算法相比取得了最好的检测结果。本文算法对复杂遥感场景的目标具有更优的检测效果。

Due to the dense arrangement and different directions of objects in remote sensing images,the existing detection algorithms are difficult to locate the instance objects accurately.Therefore,this paper proposed an omnidirectional accurate detection algorithm for dense small objects in remote sensing images.In order to improve the ability of feature extraction,the meta-ACON activation function was introduced into the residual structure of the backbone network to adaptively learn the importance of channel features.A new strengthen tconnection feature pyramid network was proposed.The lateral connection part for deep and shallow feature fusion was redesigned,and a jump connection was added between the input and output of the same level feature map to enrich the feature semantic information.The angle prediction branch was introduced,and the circular smooth label method was used to transform the angle regression problem into a classification problem.While realizing the rotation of the object frame,the problem of sudden change of the boundary of the rotation frame was solved.A post-processing method(Rotate-Soft-NMS)for rotation detection frame was designed to remove adjacent repeated rotation detection frames by suppressing the confidence of the detection frame.The experimental results on DOTA dataset show that the mAP of the proposed algorithm is 76.15%,which is 5.22%higher than the benchmark model YOLOv5m,and has achieved the best detection results compared with other advanced algorithms.The algorithm in this paper has achieved better detection results for complex remote sensing images.