Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (1): 209-217.DOI: 10.3778/j.issn.1002-8331.2107-0379

• Graphics and Image Processing • Previous Articles     Next Articles

Remote Sensing Image Aircraft Target Detection Combined with Multiple Channel Attention

LI Jie, ZHOU Shun, ZHU Xinchao, LI Yi, WANG Enguo   

  1. School of Electrical and Electronic Engineering, Hubei University of Technology, Wuhan 430068, China
  • Online:2022-01-01 Published:2022-01-06



  1. 湖北工业大学 电气与电子工程学院,武汉 430068

Abstract: Aiming at the problem of low aircraft target recognition accuracy in remote sensing images due to diversified target scales, dense targets, and poor imaging quality, a remote sensing aircraft target automatic detection model AFF-CenterNet, which combines parallel layer feature sharing structure and attention mechanism, is proposed. Firstly, the proposed method adopts an “encoding-decoding” backbone network structure to extract feature based on ResNet50. Then, the parallel layer feature sharing structure with hole convolution and attention constraints is used for feature fusion, which effectively improves the feature extraction ability of the algorithm. Finally, the quantitative comparison analysis is carried out on the public UCAS-AOD and RSOD data sets, and the detection accuracy reaches 96.78%. Compared with the faster R-CNN, SSD, YOLOv5 and the original CenterNet algorithm, the detection accuracy is improved by 6.2%, 7.2%, 1.48% and 16%, respectively. The experimental results show that the AFF-CenterNet algorithm maximizes CenterNet’s small target representation ability while maintaining a certain computational efficiency, thus improving the detection accuracy of aircraft in remote sensing images, and providing a reference for realizing rapid detection of aircraft in remote sensing images.

Key words: remote sensing aircraft image, CenterNet, attention mechanism, AFF-CenterNet

摘要: 针对尺度多样化、目标密集、成像质量较差的遥感影像上飞机目标识别精度低的问题,提出结合平行层特征共享结构和注意力机制的遥感飞机目标自动检测模型AFF-CenterNet。该方法采用“编码-解码”的主干网络结构,以ResNet50进行基础特征提取;引入空洞卷积与注意力约束的平行层特征共享结构进行特征融合,有效提高了算法的特征提取能力;在UCAS-AOD和RSOD公共遥感数据集上分别进行实验,检测精度达到96.78%,相较于Faster R-CNN、SSD、YOLOv5s和原CenterNet算法分别提高了6.2、7.2、1.48和16个百分点。实验结果表明,该AFF-CenterNet算法在保持一定计算效率的条件下最大化CenterNet的小目标表征能力,有效提升了遥感影像中飞机的检测精度,对实现遥感影像飞机快速检测具有一定的参考意义。

关键词: 遥感飞机影像, CenterNet, 注意力机制, AFF-CenterNet