Aerial military target detection algorithm based on multi-feature cross fusion and cross-layer concatenation

基于多特征交叉融合及跨层级联的航拍目标检测算法

¹ School of Computer Science and Engineering, Xi'an University of Technology, Xi'an 710021, China
² School of Weapon Science and Technology, Xi'an University of Technology, Xi'an 710021, China
³ School of Mechanical and Electrical Engineering, Xi'an University of Technology, Xi'an 710021, China

Received: 20 October 2022

Abstract

The precise detection of military targets under complex conditions is a key factor to enhance the ability of war situation generation and prediction. The current technology can not overcome the problems of smoke and occlusion interference, target height change, and different scales in aerial video. In this paper, a multi feature cross fusion and cross layer cascade aerial military target detection algorithm (YOLOv5-MFLC) is proposed. Firstly, aiming at the high confidentiality of the military targets and the shortage of battlefield aerial image resources, a real scene based aerial military target dataset is constructed, and the methods of random splicing and random extraction embedding are used for data enhancement in order to improve the diversity and generalization of targets. Secondly, aiming at the problem of complex background interference, a multi feature cross fusion attention mechanism is constructed to enhance the available information of target features. Finally, for the multi-scale problem of targets in aerial images, a cross layer cascaded multi-scale feature fusion pyramid is designed to improve the detection accuracy of cross scale targets. The experimental results show that, comparing with the existing advanced detection models, the detection accuracy of the algorithm in this paper has been greatly improved. The average accuracy of the algorithm can reach 81.0%, which is 5.2% higher than the original network. In particular, it has reached 55.9% in the smaller target category "person", which is 9.4% higher. And the experimental results further show the usefulness of the improved algorithm for small target detection. At the same time, the detection rate of this algorithm can reach 56 frame/s, which can effectively achieve accurate and fast detection of battlefield targets, and has certain experience value for guiding complex modern wars.

摘要

复杂条件下特殊目标的精确检测是增强特定场景态势生成和预测能力的关键因素。目前的技术不能克服航拍视频中出现的烟雾和遮挡干扰、目标高度变化、尺度不一等问题。因此, 提出一个多特征交叉融合及跨层级联的航拍特殊目标检测算法(YOLOv5-MFLC)。针对实际特殊目标保密性高、航拍图像资源匮乏的问题, 构建了一个基于真实场景的航拍特殊目标数据集, 并采用随机拼接和随机提取嵌入的方法进行数据增强以提高目标多样性和泛化性; 针对复杂背景干扰问题, 构建了多特征交叉融合注意力机制, 增强了目标特征的可用信息; 针对航拍图像中目标多尺度问题, 设计了跨层级联多尺度特征融合金字塔, 提高了跨尺度目标的检测准确率。实验结果表明, 与现有的先进检测模型相比, 所提算法的检测准确率有较大提升, 算法平均准确率可达到81.0%, 相比于原始网络提升了5.2%, 特别是, 在更小的目标类别"person"中达到了55.9%, 提升了9.4%, 进一步表明了所提改进算法对小目标检测的有用性。同时, 所提算法的检测速率可以达56 frame/s, 能够有效地实现实际复杂场景特殊目标的准确、快速检测, 对特殊目标的识别具有一定的指导意义。