基于旋转等变卷积的航拍红外图像目标识别算法  

Aerial Infrared Image Target Recognition Algorithm Based on Rotation Equivariant Convolution

作　　者：肖锋[1] 卢浩 张文娟 黄姝娟[1] 焦雨林 卢昭廷 李照山 XIAO Feng;LU Hao;ZHANG Wenjuan;HUANG Shujuan;JIAO Yulin;LU Zhaoting;LI Zhaoshan(School of Computer Science and Engineering,Xi’an Technological University,Xi’an 710016,Shaanxi,China;School of Ordnance Science and Technology,Xi’an Technological University,Xi’an 710016,Shaanxi,China;School of Sciences,Xi’an Technological University,Xi’an 710016,Shaanxi,China)

机构地区：[1]西安工业大学计算机科学与工程学院,陕西西安710016 [2]西安工业大学兵器科学与技术学院,陕西西安710016 [3]西安工业大学基础学院,陕西西安710016

出　　处：《兵工学报》2024年第8期2817-2827,共11页Acta Armamentarii

基　　金：国家自然科学基金面上项目(62171361);陕西省科技厅自然科学基础研究计划项目(2021JM-440);陕西省重点研发计划项目(2022GY-119)。

摘　　要：为提高传统无人机红外目标识别算法对输入图像的旋转鲁棒性,提出一种具有旋转等变性的红外图像目标识别算法。参照可见光三通道结构,将红外图像扩张为三通道以丰富输入图像的细节及边缘信息;以旋转等变卷积为基础,设计并实现能够高度保留图像旋转特征的标准旋转等变卷积模块和旋转残差模块,使得所设计模型FC-YOLOv5对图像及图像中目标旋转具有鲁棒性;加入压缩和激励注意力机制自适应地学习到每个通道的重要性,并且根据任务的需要加权调整特征图中的通道贡献,提取重要的特征信息并抑制不重要的特征信息。在航拍行人车辆数据集和海上船舶数据集上验证模型的性能,以基准模型YOLOv5s及常见轻量级目标识别任务所用模型YOLOv8s、NanoDet作为对照组模型。实验结果表明,所提算法的平均精度均值相较于基准模型能够提升2%~4%,且当输入图像具有不同角度的旋转时,能够比对照组模型识别到更多旋转目标,且识别错误更少。In order to improve the rotation robustness of the traditional UAV infrared target recognition algorithm to the input image,an infrared image target recognition algorithm with rotation equivariant is proposed.An infrared image is expanded into a three-channel image to enrich the details and edge information of the input image by referring to the visible-light three-channel structure.The standard rotation equivariant convolution FBL and rotational residual FSP modules are designed and implemented based on the rotation equivariant convolution,which can highly retain the rotation characteristics of the image,so that the FC-YOLOv5 model is robust to the rotation of the image and the target in the image;The SE attention mechanism is added to learn the importance of each channel adaptively,and the channel contribution in the feature map is weighted according to the needs of the task,so as to extract the important feature information and suppress the unimportant feature information.The performance of the model is verified on APOPV data set and SAS data set,and the benchmark model YOLOv5s and the models YOLOv8s and NanoDet used in common lightweight target recognition tasks are used as the control models.The experimental results show that the mean average precision of the proposed algorithm can be improved by 2%-4%compared to the benchmark model,and when the input image has different rotation angles,the FC-YOLOv5 can recognize more rotating targets with fewer recognition errors than those of the control model.

关 键 词：低空航拍 红外图像 多角度目标识别 旋转等变卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]