基于改进YOLOv5的空中目标部位快速分割算法  

Airborne Target Part Fast Segmentation Algorithm Based on Improved YOLOv5

作　　者：于傲泽 夏智权 魏维伟 丁霞 沈艳秋 YU Aoze;XIA Zhiquan;WEI Weiwei;DING Xia;SHEN Yanqiu(Shanghai Radio Equipment Research Institute,Shanghai 201109,China;Shanghai Engineering Research Center of Target Identification and Environment Perception,Shanghai 201109,China;Unit 93126 of People’s Liberation Army,Beijing 100038,China)

机构地区：[1]上海无线电设备研究所,上海201109 [2]上海目标识别与环境感知工程技术研究中心,上海201109 [3]中国人民解放军93126部队,北京100038

出　　处：《制导与引信》2023年第4期48-55,共8页Guidance & Fuze

基　　金：国家自然科学基金(42005110);上海市产业协同创新项目(HCXBCY-2022-059)。

摘　　要：为满足导弹武器对空中目标高效毁伤的需求,针对基于深度学习的算法对空中目标部位分割速度慢、准确率低的问题,提出一种基于改进YOLOv5的空中目标部位分割算法。将YOLOv5的检测头替换为类似YOLACT算法的分割头,使其能够完成实例分割任务;对网络进行剪枝,在不影响精度的前提下获得更快的推理速度;在Backbone网络中应用可变形卷积、增加坐标注意力机制(coordinate attention,CA)模块,丰富目标的多尺度空间及语义信息,进一步提升算法的局部特征提取能力;基于Ghost卷积设计C3_Ghost模块,替换YOLOv5原算法Neck部分的C3模块,显著降低算法计算量,并保证算法精度不受太大影响。对比与消融实验表明,所提方法与改进策略能有效降低计算量,在保证实时性的同时,算法精度也有所提升。To meet the requirements of missile weapons to destroy airborne targets efficiently,an airborne target part segmentation algorithm based on improved YOLOv5 was proposed to address the problem of slow segmentation speed and low accuracy of deep learning algorithm.The detection head of YOLOv5 was replaced with a segmentation head similar to that used in YOLACT to make it capable of completing instance segmentation tasks.The network was also pruned to obtain faster inference speed without losing accuracy.Deformable convolution and coordinate attention(CA)modules were applied to the Backbone to enrich the multi-scale spatial and semantic information of the targets and further improved the local feature extraction capability of the algorithm.In addition,the C3_Ghost module based on Ghost convolution was designed to replace the original C3 module in the Neck of YOLOv5,which significantly reduced computational cost with an acceptable level of accuracy degradation.The contrast and ablation experiments show that the proposed method and improvement strategies can effectively reduce computation cost,ensure real-time performance and improve accuracy.

关 键 词：部位分割 YOLOv5 实例分割 

分 类 号：TN957.52[电子电信—信号与信息处理]