基于CNN-LSTM的序列图像空间目标识别方法  

CNN-LSTM Based Space Object Recognition Method for Sequence Images

作　　者：齐思宇 赵慧洁 姜宏志[2,4] 李旭东[2,3,4] 王思航 郭琦[2,4] QI Siyu;ZHAO Huijie;JIANG Hongzhi;LI Xudong;WANG Sihang;GUO Qi(Institute of Artificial Intelligence,Beihang University,Beijing 100191,China;School of Instrumentation and Optoelectronic Engineering,Beihang University,Beijing 100191,China;Qingdao Research Institute of Beihang University,Qingdao 266104,Shandong,China;Key Laboratory of Precision Opto-Mechatronics Technology,Ministry of Education,Beijing 100191,China)

机构地区：[1]北京航空航天大学人工智能研究院,北京100191 [2]北京航空航天大学仪器科学与光电工程学院,北京100191 [3]北京航空航天大学青岛研究院,青岛266104 [4]教育部精密光机电一体化技术重点实验室,北京100191

出　　处：《上海航天(中英文)》2025年第2期186-193,共8页Aerospace Shanghai(Chinese&English)

基　　金：航天科技集团应用创新资助项目(6230109004)。

摘　　要：针对现有的基于序列图像的空间目标识别方法难以在特征层级进行融合的问题,提出了将深度卷积网络(CNN)与循环神经网络(RNN)相结合的方法,并对网络模型加以改进。针对单幅图像如何作为序列特征输入的问题,对卷积网络的末端进行修改,将特征图作为序列特征输入;针对序列特征如何映射到目标类别的问题,对长短期记忆网络(LSTM)网络末端进行修改,增加了新的全连接层,得到输出类别。使用0.001~0.006高斯噪声水平训练,以0.007~0.010作为测试集,识别平均准确率(mAP)由90.7%提升至99.16%;训练集与测试集在不同姿态情况下,mAP为94.71%。网络参数量仅为283.0 M。现有的仅在结果层级融合进行识别的问题得到了有效解决。To address the challenge of feature-level fusion in sequence image-based space target recognition,this paper proposes a method combining convolutional neural networks(CNNs)and recurrent neural networks(RNNs)with model improvements.In view of the problem of how to use a single image as a sequence feature input,the CNN is modified,in which the feature maps are used as the sequential inputs.In view of the problem of how the sequence features map to the target categories,the long short-term memory(LSTM)network is modified,in which the output layer is enhanced with a new fully connected layer to predict the target categories.Training with the Gaussian noise levels of 0.001~0.006 and testing at 0.007~0.010 achieve a mean average precision(mAP)improvement from 90.7%to 99.16%.Under different postural conditions,the mAP reaches 94.71%.The model has only 283.0 M parameters,effectively addressing the limitations of result-level fusion in existing methods.

关 键 词：目标识别 序列图像 空间目标 卷积网络(CNN) 循环神经网络(RNN) 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]