基于改进YOLOX的遥感图像目标检测

Remote sensing image target detection based on improved YOLOX

为解决遥感图像目标检测中存在的密集小目标易被漏检以及结构复杂目标检测精度低等问题,设计了一种基于YOLOX的检测器.首先,构建一种基于全局平均池化金字塔结构的注意力网络(convolutional block spatial net,CBSPNet),并将CBSPNet插入锚点优化空洞卷积改进的残差块内,以增强特征的提取能力和优化对目标边缘的检测;其次,在主干网络中新增一个浅层的输出尺度,使得网络对小目标更为敏感;最后,引入跳跃连接构建优化的特征融合网络,提高特征在多尺度空间中的融合能力.实验结果表明:改进的YOLOX模型(R-YOLOX)对遥感密集小目标和不规则大目标均具有良好的检测性能,在多场景的检测任务中展现了较强的鲁棒性;R-YOLOX网络在RSOD遥感图像数据集上的平均检测精度较YOLOX提高了1.15%,且检测速率达31帧·s^(-1),满足实时检测的要求.

In order to solve the problems of dense small targets easy to miss detection,complex target structure and low detection accuracy in remote sensing image target detection,a YOLOX-based detector is designed.Firstly,an attention network based on global average pooled pyramid structure(CBSP-Net)is constructed,and CBSP-Net is inserted into the residual block improved by anchor optimization hole convolution to enhance the extraction ability of features and optimize the detection of target edges.Secondly,a shallow output scale is added to the backbone network to make the network more sensitive to small targets.Finally,the hopping connection is introduced to construct an optimized feature fusion network to improve the feature fusion ability in multi-scale space.Experimental results show that the improved YOLOX model(R-YOLOX)has good detection performance for both dense small targets and irregular large targets in remote sensing,and shows strong robustness in multi-scene detection tasks.The average detection accuracy of R-YOLOX network on RSOD dataset is 1.15%higher than that of YOLOX network,and the detection rate reaches 31 frames·s^(-1),which meets the requirements of real-time detection.