基于小样本集的SAR图像船只稳健表征学习
Learning of Ship Robustness Characterization in SAR Images Based on Small Sample Sets

随着合成孔径雷达(Synthetic Aperture Radar, SAR)系统在军事及民用领域的广泛使用，SAR图像数据的规模开始急速扩增，不同应用场景下的SAR图像目标分类需求也不断增多。传统目标分类算法需要在分析数据集后针对图像特性手动设计特征提取器，设计过程复杂繁琐且专业知识依赖性强，难以满足实际需求，因此深度学习方法开始被引入SAR图像目标分类领域。但是，由于现有带标签SAR图像数据集规模较小，且SAR图像与光学图像在图像特征上存在一定差异，直接将光学图像卷积神经网络模型应用到SAR图像上往往很难取得理想效果。针对上述问题，本文提出了基于卷积神经网络迁移学习的SAR图像目标分类方法，相对于SAR图像而言，光学图像数据的来源更为广泛，获取也更加容易。通过迁移学习的方式，利用在大规模光学数据集上充分训练的预训练模型来辅助SAR图像分类模型的训练，从而实现对船只进行的有效分类识别。
With the widespread use of Synthetic Aperture Radar (SAR) systems in the military and civil fields, the scale of SAR image data begins to expand rapidly, and the demand for SAR image target classi-fication in different application scenarios also increases. Traditional target classification algorithms need to manually design feature extractors for image characteristics after analyzing data sets. The design process is complex and tedious, and the professional knowledge is highly dependent, so it is difficult to meet the actual requirements. Therefore, deep learning method has been introduced into the SAR image target classification field. However, due to the small scale of the existing SAR image data set and the certain difference in image characteristics between SAR image and optical image, it is often difficult to obtain ideal results when the optical image convolutional neural net-work model is directly applied to SAR image. To solve the above problems, this paper proposes a method for SAR image target classification based on convolutional neural network migration learning. Compared with SAR images, optical image data are more widely sourced and easier to obtain. By means of transfer learning, the pretraining model fully trained on large-scale optical data set is used to assist the training of SAR image classification model, so as to realize the effective classification and recognition of ships.