基于检索任务和改进ResNet的机场跑道异物目标识别方法  

Foreign object debris target recognition method for airport runways based on retrieval task and improved ResNet

作　　者：刘广 李海旭 侯腾 王晓成 LIU Guang;LI Haixu;HOU Teng;WANG Xiaocheng(System Engineering Research Institute,China State Shipbuilding Corporation Limited,Beijing 100094,China)

机构地区：[1]中国船舶集团有限公司系统工程研究院,北京100094

出　　处：《计算机应用》2024年第S01期309-313,共5页journal of Computer Applications

摘　　要：针对机场跑道异物(FOD)普遍存在目标尺度差异大、小目标占比大且同种材质目标形态不统一等特性,提出一种基于检索任务和改进ResNet的FOD目标识别方法。首先,在网络训练阶段通过训练得到特征提取网络,将训练集中的样本数据输入特征提取网络,得到样本特征库;其次,在测试阶段,将测试样本输入特征提取网络,获取测试样本特征向量,计算测试样本特征向量与样本特征库中特征向量的距离;最后,将距离最近的样本类型赋予该测试样本。所提方法采用的检索任务思想在出现新类型目标时,无需重新训练网络模型,大幅地节约了人力和物力成本。另外,所提方法将ResNet50作为特征提取网络,并在此基础上增加多尺度特征提取模块(MSM)、通道注意力模块(CAM)和特征融合模块。实验结果表明,与Faster-RCNN和YOLO-V5-n方法相比,所提方法的平均识别率分别提高了2.44和0.18个百分点,能在实际应用中实现更精确的FOD识别。To address challenges inherent in Foreign Object Debris(FOD)detection on airport runways,such as large difference in target scale,high proportion of small targets and diverse morphologies within the same material category,a FOD target recognition method was proposed based on retrieval task and an enhanced ResNet architecture.During the network training phase,a specialized feature extraction network was trained to construct a feature library using sample data from the training set.In the testing phase,est samples were input into the feature extraction network for generating their feature vectors.The distances between the feature vector of a test sample and those stored in the sample feature library were calculated,and the sample type exhibiting the closest proximity was assigned to the test sample.The methodology,rooted in retrieval task principles,representd a significant advancement by eliminating the need for laborious network retraining when confronted with new target types.This innovation resulted in substantial reduction in human and material costs.Furthermore,ResNet50 was used as the feature extraction network,enriched with specialized modules including Multi-Scale Feature Extraction Module(MSM),Channel Attention Module(CAM),and Feature Fusion Module.Experimental results show the proposed method has the improvements of 2.44 and 0.18 percentage points in average recognition rate compared to Faster-RCNN and YOLO-V5-n methods,respectively.It can realize more accurate FOD recognition in practical airport runway applications.

关 键 词：机场跑道异物 检索任务 目标识别 多尺度特征提取 注意力模块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]