基于残差密集连接注意力的红外目标识别算法  

An Infrared Target Recognition Algorithm Based on Residual Dense Attention

作　　者：刘琦妍 张凯[1] 王田田 杨尧[1] LIU Qiyan;ZHANG Kai;WANG Tiantian;YANG Yao(Unmanned Systems Research Institute,Northwestern Polytechnical University,Xi'an 710000,China)

机构地区：[1]西北工业大学无人系统技术研究院,西安710000

出　　处：《电光与控制》2024年第6期31-35,共5页Electronics Optics & Control

摘　　要：针对红外目标识别时目标与干扰混淆、干扰遮挡目标导致算法性能下降的问题,提出了一种基于残差密集连接注意力的空中红外目标识别算法。首先,为跨层实现浅层与深层特征的融合、获取融合深度特征、提升特征复用性并增强语义信息,提出了改进的残差密集连接模块;其次,为增强融合深度特征的自适应表达能力,设计了并行混合注意力模块;最后,通过在大量红外数据集上的测试证明,相比于GoogLeNet算法,所提算法平均识别正确率提高了1.9个百分点,证明了算法的有效性。To address the problems as degradation of algorithm performance because the target is blocked by interference and confusion of target and interference in infrared target recognition,a new airborne infrared target recognition algorithm is proposed based on residual dense connection attention.Firstly,to fuse shallow and deep features across layers,obtain fusion depth features,enhance feature reuse performance and strengthen semantic information,an improved residual dense block is proposed.Secondly,to enhance the adaptive expression ability of the fused depth features,a parallel mixed attention block is designed.Finally,the test on a large quantity of infrared datasets shows that the algorithm's average recognition accuracy is increased by 1.9 percentage points compared with that of the GoogLeNet algorithm,which proves the validity of the algorithm.

关 键 词：红外空空导弹 目标识别 特征融合 GoogLeNet 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]