基于多尺度特征增强的遥感图像目标检测方法
Remote Sensing Image Object Detection Algorithm Based on Multi-Scale Feature Enhancement

针对遥感图像目标检测中存在的背景复杂、目标像素数少以及目标尺度变化大等问题，本文提出一种基于多尺度特征增强的遥感图像目标检测方法。首先，使用具有高分辨率输出的HRNet网络替换ResNet作为主干网络，强化对遥感目标位置信息的提取；其次，在HRNet中引入注意力机制，抑制复杂背景噪声的干扰；最后，设计多尺度特征增强金字塔网络，进一步增强网络的多尺度特征信息表达。实验结果表明，相较于原始Cascade R-CNN目标检测方法，所提方法的目标检测均值平均精度提高了5.32%；在与经典目标检测方法的对比实验中，所提方法也表现出较好的检测性能。
To address the problems of complex image background, small number of object pixels and large variation of object scale in remote sensing image object detection, we propose a remote sensing image object detection method based on multi-scale feature enhancement. First, the HRNet network with high-resolution output is used to replace ResNet to strengthen the backbone network to obtain the location of remote sensing objects; second, the attention mechanism is introduced into HRNet to suppress the interference of complex background noise; finally, the multi-scale feature-enhanced pyramid network is designed to further enhance the multi-scale information representation of the pyramid network. The results of the experiment show that compared with the Cascade R-CNN object detection method, the mean accuracy of the proposed method is improved by 5.32%, and the proposed method also shows better detection performance in comparison with the classical object detection method.