卷积神经网络的出现使得深度学习在视觉领域取得了巨大的成功,并逐渐延伸到合成孔径雷达(SAR)图像识别领域。然而,SAR图像样本量不足,难以支撑卷积神经网络的训练需求,并且SAR图像包含大量相干斑噪声及不确定性,网络结构的设计较为困难。所以,深度学习在SAR图像识别领域的应用受到阻碍。针对上述问题,文中提出一种基于数据扩维的SAR目标识别性能提升方法,通过对原始SAR 图像进行相关预处理操作并把处理后图像与原始图像结合,从而将一维的原始数据扩充成多维数据来作为训练样本。该扩维方法不仅间接扩充了样本量来支撑网络训练,同时也在网络训练前加入了“主动学习冶影响,所以无需针对SAR图像特性来构建复杂卷积网络,而采用成熟、简单的网络进行训练就可以达到理想的测试精度。最后,使用MSTAR 数据对该方法进行了性能验证,实验结果显示了所提方法的有效性。

The emergence of convolutional neural networks has made deep learning a great success in the field of computer vision, and has been gradually extended to the field of synthetic aperture radar (SAR) image recognition. However, the lack of sample size of SAR images makes it difficult to support the training needs of convolutional neural networks. Moreover, SAR images contain a large amount of speckle noise and uncertainty. The design of the network structure also becomes a difficulty. Therefore, deep learning is difficult to be applied in the field of SAR image recognition. Aiming at this issue, a method for improving performance of SAR target recognition based on data expansion is presented. Several SAR samples generated by image preprocessing are aggreated so that the one-dimensional data is extended to multidimensional data. It is unnecessary to construct a complex convolutional network based on the characteristics of the SAR image, and only a mature and simple network is used for training to achieve the desired test accuracy. In the end, the performance of the proposed method is validated based on the MSTAR database, and experimental results show the effectiveness of the proposed method.

TN957.51

中国博士后科学基金面上资助项目