基于改进FCOS网络的光学遥感影像目标检测算法

针对当前遥感影像背景复杂、目标尺度小等情况导致的检测精度偏低的问题，基于FCOS网络提出了一种结合位置注意力和感受野增强的遥感影像目标检测算法PARF-FCOS；该算法构造了一种位置注意力模块，并利用该模块对特征提取网络进行改进，增强网络对目标信息的提取能力；在特征融合阶段使用感受野模块（RFB, receptive field block）增强浅层特征图，利用目标上下文信息进行辅助判断，提升网络对小尺度目标的检测能力；在训练过程中，引入距离交并比损失（DIoU loss，distance intersection over union loss）进行边界框回归，通过优化目标框与预测框中心点之间的距离，使回归过程更加平稳和准确;在公开数据集DIOR上评估了PARF-FCOS目标检测算法，实验结果表明，相较于原始FCOS，算法的平均精确度均值提高了4.3%，达到70.4%，检测速度达到23.2FPS。

Object Detection Algorithm of Optical Remote Sensing Imagery Based on An Improved FCOS Network

Aiming at the problem of low detection accuracy caused by complex background and small objects in remote sensing imageries, a remote sensing imagery object detection algorithm PARF-FCOS based on FCOS network is proposed. The algorithm constructs a position attention module, and uses the module to reconstruct the feature extraction network to enhance the ability of the network to extract target information; In the feature fusion stage, RFB(receptive field block) is used to enhance the shallow feature map, and the target context information is used for auxiliary judgment to improve the detection ability of the network for small-scale objects; During training, DIOU loss(distance intersection over union loss) is introduced for boundary box regression. By optimizing the distance between the center point of the target box and the prediction box, the regression process is more stable and accurate. Experiments are carried out with public dataset DIOR. Compared with the original FCOS, the mean Average Precision of the algorithm is improved by 4.3%, up to 70.4%, and the detection speed is 23.2 FPS.