一种基于AlexNet迁移学习模型的空间目标ISAR像识别方法

Recognition method of space target ISAR images based on AlexNet transfer learning model

利用ISAR像进行快速有效的空间目标识别是目前空间态势感知领域研究的热点,然而受到样本数据量的限制,直接训练深度神经网络容易产生过拟合,难以准确预测;而传统基于手动提取ISAR像特征的目标识别方法,操作繁琐且需要耗费大量的时间与精力。针对上述问题,提出一种基于迁移学习快速自主识别空间目标的新方法,该方法通过角度旋转、方位向距离向尺度变换、斑点噪声注入等方式对小样本空间目标ISAR像数据集增强;以AlexNet预训练模型为基础,对模型的“深”、“浅”层分别设置差异化的学习率;通过反向传播方式对AlexNet模型的权值进行微调从而实现模型迁移。训练测试结果表明,数据增强方法可以有效提高模型的分类性能,该方法可以实现小样本数据集下目标的自主快速识别,与传统方法相比具有更好的分类性能。

To solve the problem of fast and autonomous recognition of space targets using Inverse Synthetic Aperture Radar(ISAR)images under the condition of small-scale data sets,a recognition method based on data augmentation(DA)and transfer learning(TL)was proposed.Starting with ISAR imaging of space target,the influence of image projection plane(IPP),azimuth resolution,sampling rate and other factors on ISAR imaging was analyzed,and then DA methods including rotation,scaling in azimuth and cross-range direction,and speckle noise injection was proposed.Then the DA method was used to augment the small-scale ISAR image data set of space targets.Based on the pre-trained AlexNet model,differentiated learning rates were set for the“deep”and“shallow”layers of the model,and the weights of AlexNet model were fine-tuned by backpropagation.On this basis,AlexNet transfer learning model(ATLM)was established.The training and testing results of ATLM under four TL settings on ISAR image datasets of 5 kinds of space targets show that ATLM can realize target recognition autonomously and quickly under the condition of small-scale data sets,and has better classification performance than traditional methods.DA method can effectively improve the classification performance of the ATLM.