基于改进YOLO的装甲目标毁伤状态估计方法OA
Damage State Estimation Method for Armored Targets Based on Improved YOLO
针对战场上装甲目标毁伤状态检测所面临的图片背景复杂、毁伤程度各异及毁伤特征多变等挑战,提出一种轻量级装甲目标毁伤状态评估方法——挤压-激励(squeeze-and-excitation,SE)与特征融合增强的 YOLO 模型.首先,在 Focus 模块中添加 SE 注意力机制,设计了 SE-Focus 特征提取模块,通过增强关键特征和减少信息冗余来提升模型在复杂环境下的适应性和检测效果;其次,融合加权跨尺度双向特征金字塔网络(bi-directional feature pyra-mid network,BiFPN)的多尺度特征融合优势与深度可分离卷积(depthwise separable convolution,DSConv)的高效计算特性,在保留 BiFPN 的优异跨尺度特征交互能力的同时,借助 DSConv 的特性降低模型复杂度与运算成本;最后,引入 SIoU 损失函数,增强模型的鲁棒性及训练稳定性.在装甲目标毁伤坦克数据集上的实验结果表明,参数量不变的情况下,所提模型相比原模型 mAP50 提高 1.8 个百分点,召回率提高 1.9 个百分点,mAP50-95 提高 1.7 个百分点,验证了改进策略的有效性.
A lightweight armored target damage state evaluation method,the squeeze-and-excitation(SE)and feature fusion enhanced YOLO model,was proposed to address the challenges caused by com-plex battlefield image backgrounds,varying damage levels,and dynamic damage characteristics in target damage state assessment.Firstly,a SE-Focus feature extraction module was designed by adding SE atten-tion mechanism in the Focus module.By enhancing key features and reducing information redundancy,the adaptability and detection performance of the model in complex environments were improved.Second-ly,the advantages of multi-scale feature fusion from the weighted bi-directional feature pyramid network(BiFPN)were integrated with the efficient computing characteristics of depthwise separable convolution(DSConv).While retaining the excellent cross scale feature interaction ability of BiFPN,the feature of DSConv was utilized to reduce model complexity and computational cost.Finally,the SIoU loss function was introduced to enhance the model's robustness and training stability.Experimental results demonstra-ted that when tested on the damaged tank dataset of armored targets,a 1.8 percentage points improve-ment in mAP50,a 1.9 percentage points increase in recall rate,and a 1.7 percentage points rise in mAP50-95 were achieved by the proposed model compared to the original model while the same parameter quantity was maintained,which verified the effectiveness of the proposed improvements.
宋轩;李新家;徐旺旺;佘维
郑州大学 网络空间安全学院 河南 郑州 450002郑州大学 网络空间安全学院 河南 郑州 450002郑州大学 网络空间安全学院 河南 郑州 450002郑州大学 网络空间安全学院 河南 郑州 450002||郑州市区块链与数据智能重点实验室 河南 郑州 450002
信息技术与安全科学
毁伤检测YOLO模型SIoU损失函数SE注意力机制双向特征金字塔网络
damage detectionYOLO modelSIoU loss functionSE attention mechanismBiFPN
《郑州大学学报(理学版)》 2026 (5)
14-21,8
目标易损性评估全国重点实验室开放基金项目(YSX2024KFPG004)
评论