测绘学报 ›› 2024, Vol. 53 ›› Issue (7): 1384-1400.doi: 10.11947/j.AGCS.2024.20230455

• 摄影测量与遥感 • 上一篇    下一篇

遥感影像目标检测多尺度熵神经网络架构搜索

杨军1,2(), 解恒静1, 范红超3, 闫浩文1   

  1. 1.兰州交通大学测绘与地理信息学院,甘肃 兰州 730070
    2.兰州交通大学电子与信息工程学院,甘肃 兰州 730070
    3.挪威科技大学土木与环境工程学院,挪威 特隆赫姆 7491
  • 收稿日期:2023-10-07 发布日期:2024-08-12
  • 作者简介:杨军(1973—),男,博士,教授,博士生导师,主要研究方向为三维模型空间分析、遥感大数据智能解译、深度学习。E-mail:yangj@mail.lzjtu.cn
  • 基金资助:
    国家自然科学基金(42261067)

Multi-scale entropy neural architecture search for object detection in remote sensing images

Jun YANG1,2(), Hengjing XIE1, Hongchao FAN3, Haowen YAN1   

  1. 1.Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou 730070, China
    2.School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
    3.Department of Civil and Environmental Engineering, Norwegian University of Science and Technology, Trondheim 7491, Norway
  • Received:2023-10-07 Published:2024-08-12
  • About author:YANG Jun (1973—), male, PhD, professor, PhD supervisor, majors in 3D model spatial analysis, intelligent interpretation of remotely sensed big data, deep learning. E-mail: yangj@mail.lzjtu.cn
  • Supported by:
    The National Natural Science Foundation of China(42261067)

摘要:

针对传统神经网络架构搜索需要耗费大量时间用于超网训练,搜索效率较低,搜索得到的模型无法高效解决遥感影像中多尺度目标检测困难、背景复杂度高的问题,本文提出采用多尺度熵神经网络架构搜索的方法进行遥感影像目标检测。首先,在搜索空间的基础模块中加入特征分离卷积以代替残差模块中的常规卷积,减少遥感影像中由于背景复杂度高而造成的信息间干扰,提高网络模型在复杂背景下的检测性能;然后,引入最大熵原理,计算搜索空间中每个候选网络的多尺度熵,将多尺度熵与特征金字塔网络相结合,以兼顾遥感影像大、中、小目标的检测;最后,在不进行参数训练的情况下利用渐进式进化算法搜索得到多尺度熵最大的网络模型用于目标检测任务,在保证模型检测精度的同时,提升网络搜索效率。本文方法在RSOD、DIOR和DOTA数据集上的平均检测精度均值分别达到93.1%、75.5%和73.6%,网络搜索时间为8.1 h。试验结果表明,与当前基准方法相比,本文方法能够显著提升网络的搜索效率,在目标检测任务中更好地结合了不同尺度下的特征并解决了影像背景复杂度高的问题。

关键词: 遥感影像, 神经网络架构搜索, 目标检测, 特征分离卷积, 最大熵, 多尺度熵, 渐进式进化

Abstract:

Aiming at the traditional neural architecture search requires an enormous amount of time for supernet training, search efficiency is suboptimal, and the searched model can not efficiently solve the problem of multi-scale object detection difficulty and high background complexity in remote sensing images. This paper proposes a multi-scale entropy neural architecture search method for object detection in remote sensing images. At first, the feature separation convolution is added to the base block of the search space instead of the regular convolution in the residual block, which reduces the information redundancy in remote sensing images due to the high background complexity, and improves the detection performance of the network modeluner the complex background. Next, the maximum entropy principle is introduced to calculate the multi-scale entropy of each candidate network in the search space, and the multi-scale entropy is combined with the feature pyramid network to balance the detection of large, medium and small objects in remote sensing images. Finally, the network model with maximum multi-scale entropy is obtained by searching without parameter training using progressive evolutionary algorithm for the object detection task. The model ensures detection accuracy while improving the search efficiency. The proposed algorithm achieves a mean average precision of 93.1%, 75.5% and 73.6% on the RSOD, DIOR and DOTA datasets, respectively, with a network search time of 8.1 hours. The experimental results demonstrate that the proposed algorithm can significantly improve the search efficiency of the network, combine the features at different scales better and solve the problem of high image background complexity in the object detection task compared with the current benchmark methods.

Key words: remote sensing, neural architecture search, object detection, feature separation convolution, maximum entropy, multi-scale entropy, progressive evolution

中图分类号: