计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (15): 210-219.DOI: 10.3778/j.issn.1002-8331.2110-0185

• 模式识别与人工智能 • 上一篇    下一篇

改进LDS_YOLO网络的遥感飞机检测算法研究

吴杰,高策,余毅,张艳超,裴玉,马少峰   

  1. 1.中国科学院 长春光学精密机械与物理研究所,长春 130033
    2.西昌卫星发射中心,四川 西昌 615099
  • 出版日期:2022-08-01 发布日期:2022-08-01

Research on Improved LDS_YOLO Network Remote Sensing Aircraft Detection Algorithm

WU Jie, GAO Ce, YU Yi, ZHANG Yanchao, PEI Yu, MA Shaofeng   

  1. 1.Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, China
    2.Xichang Satellite Launch Center, Xichang, Sichuan 615099, China
  • Online:2022-08-01 Published:2022-08-01

摘要: 为解决遥感飞机检测算法网络计算复杂、检测精度低的问题,以主流网络YOLOv4为基础,从提高精度和简化模型两个方面进行改进研究,提出一种轻量级多尺度监督网络LDS_YOLO(light dense supervision YOLO)。针对遥感飞机目标细节信息提取不足的问题,改进三组多尺度融合预测层结构,在每一个支路第一次上采样前的四个卷积块之间设计密集连接方式,可以增强融合不同尺度飞机,丰富特征细节信息,提高预测准确率;针对目标特征关联度低的问题,引入一致性监督损失函数,通过监督分类网络辅助预测的同时提高检测精度;通过增加包含全局平均池化层、全连接层和特征映射层的轻量化模块,调整通道结构减少权重模型的特征冗余,降低网络参数量。在保证检测率的基础上将模型参数量降低为3.6×106,计算量为77?MFLOPs,测试检测率比原始模型损失不到2.3%,速度达到17?frame/s;通过与主流检测算法进行对比,分析轻量化后算法模型的抗过拟合能力和鲁棒性,证明轻量化遥感飞机目标检测算法的有效性和可行性。

Abstract: In order to solve the problems of complex network calculation and low detection accuracy of remote sensing aircraft detection algorithm, this paper improves the research from two aspects of improving accuracy and simplifying model based on the mainstream network YOLOv4, and proposes a lightweight multi-scale monitoring network LDS_YOLO(light dense supervision YOLO). Aiming at the problem of insufficient extraction of remote sensing aircraft target details, the structure of three groups of multi-scale fusion prediction layers is improved, and the dense connection mode is designed between the four convolution blocks before the first up-sampling of each branch, which can enhance the fusion of aircraft of different scales, enrich the feature details and improve the prediction accuracy. Aiming at the problem of low correlation degree of target features, the consistency supervision loss function is introduced to improve the detection accuracy while assisting prediction by supervision classification network. By adding lightweight modules containing global average pooling layer, full connection layer and feature mapping layer, the channel structure is adjusted to reduce feature redundancy of the weight model and reduce the number of network parameters. On the basis of ensuring the detection rate, the number of model parameters is reduced to 3.6×106, and the computational cost is 77?MFLOPs. The test detection rate loss is less than 2.3% compared with the original model, and the speed is up to 17 frame/s. The anti-over-fitting ability and robustness of the lightweight algorithm model are analyzed by comparing with the mainstream detection algorithm, and the validity and feasibility of the lightweight remote sensing aircraft target detection algorithm are proved.