计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (22): 247-256.DOI: 10.3778/j.issn.1002-8331.2104-0213

• 图形图像处理 • 上一篇    下一篇

改进YOLOv4的密集遥感目标检测

谢俊章,彭辉,唐健峰,侯奕辰,曾庆喜   

  1. 成都信息工程大学 软件工程学院,成都 610225
  • 出版日期:2021-11-15 发布日期:2021-11-16

Improved YOLOv4 for Dense Remote Sensing Target Detection

XIE Junzhang, PENG Hui, TANG Jianfeng, HOU Yichen, ZENG Qingxi   

  1. College of Software Engineering, Chengdu University of Information Technology, Chengdu 610225, China
  • Online:2021-11-15 Published:2021-11-16

摘要:

基于遥感目标在密集分布和背景复杂场景中因特征提取和表达能力的不足而存在漏检和检测效果不佳的问题,提出了改进YOLOv4的遥感目标检测算法。对用于检测目标的锚框(anchor)用[K]-means聚类算法重新聚类来减少网络计算量;改进特征提取网络结构,引入残差连接取缔网络中连续卷积操作来提高密集目标特征提取能力;在特征提取网络中激活函数加入自适应激活与否的特征激活平滑因子,而在PANet特征融合网络结构中采用Mish激活函数,增强网络对非线性特征的提取能力,从而提升网络的特征提取能力,提高遥感目标在密集分布场景中的检测效果。将所提算法和原始的YOLOv4目标检测算法在遥感图像数据集上进行对比实验,改进YOLOv4算法在实验选用的遥感图像测试数据集上的平均准确率均值(mAP)达到85.05%,与YOLOv4算法相比,mAP提升了5.77个百分点。实验结果表明,在单目标密集分布和多目标混合分布等背景复杂条件下,改进YOLOv4算法具有更好的检测效果。

关键词: 遥感图像, 目标检测, YOLOv4, 网络结构, 激活函数

Abstract:

Based on the problems of missed detection and poor detection effect of remote sensing targets in densely distributed and complex background scenes due to insufficient feature extraction and expression capabilities, an improved YOLOv4 remote sensing target detection algorithm is proposed. The anchor uses the [K]-means clustering algorithm to re-cluster for reducing the amount of network calculation; improve the feature extraction network structure, introduce residual connections to eliminate continuous convolution operations in the network to improve the feature extraction capabilities of dense targets; in the feature extraction network, the activation function adds an adaptive feature activation smoothing factor, and the Mish activation function is used in the PANet feature fusion network structure to enhance the network’s ability of extracting nonlinear features, enhance the network’s feature extraction ability, and improve the denseness of remote sensing targets and the detection effect. The proposed algorithm and the original YOLOv4 target detection algorithm are compared on the remote sensing image data set. The improved YOLOv4 algorithm has an average accuracy rate(mAP) of 85.05% on the remote sensing image data set selected in the experiment, which is similar to the YOLOv4 algorithm. Compared with that, mAP is increased by 5.77 percentage points. The experimental results show that the improved YOLOv4 algorithm has better performance under complex background conditions such as single target dense distribution and multi-target mixed distribution.

Key words: remote sensing image, target detection, YOLOv4, network structure, activation function