融合空域和频域网络模型的SAR图像识别

A Hybrid Model of Space and Frequency Domain for SAR Image Recognition

深度学习方法已经被广泛应用于合成孔径雷达(SAR)图像识别,并取得了不错的效果。但是大多数深度学习方法仅提取空间域的图像信息,未考虑SAR图像的频域散射特性,丢失了部分关键特征。为解决上述问题,文中提出了基于融合空间域和频域网络模型的SAR图像识别的端到端的深度学习框架。首先,将原始空间域图像通过频域变换方法转换为频域图像,对频域图像进行信道选择获取有效频域信号,将有效频域信号输入到频域主干网络提取频域特征;然后,将原始空间域图像输入到空间域主干网络,提取空间域特征;最后,通过网络模型融合空间域特征和频域特征,充分利用SAR图像的空间域像素信息和频域散射特性,进一步提取出目标的本质特征。文中所提方法在公共数据集MSTAR上进行了大量的实验,验证了模型的有效性和鲁棒性。

Deep learning methods have been widely used in synthetic aperture radar(SAR)image recognition,and have achieved good results.However,most deep learning methods only extract information of image in the spatial domain,without considering the frequency domain characteristics of SAR images,and some effective features are lost.In order to solve the above-mentioned problems,a hybrid model of space and frequency domain for SAR image recognition,which is an end-to-end deep learning framework is proposed.First,the image of spatial domain is converted into a signal of frequency domain through frequency domain transformation method.The image channel selection of frequency domain is performed to obtain effective frequency domain signals,and the effective frequency domain signals are input to the frequency-domain-based backbone network to extract features of frequency domain.Then,the image of spatial domain is input to the spatial-domain-based backbone network to extract the features of spatial domain.Finally,through the fusion network to merge spatial domain features and frequency domain features,it makes full use of the spatial domain pixel information and frequency domain characteristics of the SAR image,and further extracts the essential features of the target.The method proposed in this paper has been tested by a large number of experiments on the public dataset MSTAR to verify the effectiveness and robustness of the model.