Journal of Frontiers of Computer Science and Technology ›› 2024, Vol. 18 ›› Issue (6): 1543-1555.DOI: 10.3778/j.issn.1673-9418.2305090

• Graphics·Image • Previous Articles     Next Articles

Remote Sensing Image Object Detection Algorithm Based on Multi-branch Feature Mapping

MIN Feng, KUANG Yonggang, HAO Linlin, PENG Weiming   

  1. Hubei Province Key Laboratory of Intelligent Robot, School of Computer Science and Engineering, Wuhan Institute of Technology, Wuhan 430205, China
  • Online:2024-06-01 Published:2024-05-31

多分支特征映射的遥感图像目标检测算法

闵锋,况永刚,郝琳琳,彭伟明   

  1. 武汉工程大学 计算机科学与工程学院 智能机器人湖北省重点实验室,武汉 430205

Abstract: Due to the complex background, small and dense targets, and large scale continuous changes in remote sensing images, universal object detectors are difficult to adapt well, resulting in poor detection performance. To address the above issues, a multi-branch feature mapping based remote sensing image object detection algorithm is proposed based on the YOLOv5s model. Firstly, a RepVGG module combined with gated channel transformation is designed using structural reparameterization technology. Its series structure is used to replace the C3 module of the original backbone network, aggregating global contextual information and enhancing feature expression and extraction capabilities. Secondly, the adaptive exponential weighted pooling method and the sampling method of inverse process reconstruction feature fusion network are used to maximize the retention of feature information and improve the detection performance of smaller targets. Finally, the combination of EIoU and Focal Loss is introduced as the loss function of the model to optimize the regression rate of the prediction box and the loss weight distribution of difficult and easy samples, further improving the positioning accuracy. The experimental results on the DIOR and NWPU VHR-10 datasets show that the average accuracy of the proposed algorithm reaches 92.2% and 92.5%, respectively, which is 3.5 percentage points and 5.6 percentage points higher than YOLOv5s, achieving better detection performance. At the same time, the real-time performance also meets the requirements of remote sensing image object detection in actual scenes.

Key words: remote sensing images, structural reparameterization, gated channel transformation, sampling method, loss weight allocation

摘要: 由于遥感图像具有背景复杂、目标较小且密集、尺度连续变化大等特点,通用目标检测器难以较好地适应,导致检测效果不佳。针对以上问题,基于YOLOv5s模型,提出一种多分支特征映射的遥感图像目标检测算法。首先,利用结构重参数化技术设计一种结合门控通道转换的RepVGG模块,采用其串联结构替换原主干网络的C3模块,聚合全局上下文信息,增强特征表达和特征提取能力;其次,使用自适应指数加权池化方法以及逆过程重构特征融合网络的采样方式,最大化地保留特征信息,改善较小目标的检测效果;最后,引入EIoU和Focal Loss组合作为模型的损失函数,优化预测框的回归速率以及难易样本的损失权重分配,进一步提高定位精度。在DIOR和NWPU VHR-10数据集上的实验结果表明,提出算法的平均精度均值分别达到92.2%、92.5%,较YOLOv5s分别提高了3.5个百分点、5.6个百分点,达到了更好的检测效果,同时实时性也满足实际场景下的遥感图像目标检测。

关键词: 遥感图像, 结构重参数化, 门控通道转换, 采样方式, 损失权重分配