loading page

Improved SAR feature fusion with convolutional neural networks and moment methods
  • Chunqain He,
  • Dongsheng Li,
  • Yang Gao
Chunqain He
National University of Defense Technology

Corresponding Author:chunqianhe@nudt.edu.cn

Author Profile
Dongsheng Li
National University of Defense Technology
Author Profile
Yang Gao
Air Force Engineering University
Author Profile


Synthetic aperture radar (SAR) is an important means of obtaining battlefield information. Correct identification of SAR images is essential. Hence, we propose a new method of SAR image recognition based on multi-feature fusion. Convolutional neural networks (CNNs) are based on local pixels and fuse this information at deep layers to obtain deep features. The moment method focuses on the entire image and obtains global moment features of different orders. We fuse the two feature types and choose a suitable feature fusion. First, we propose a Q_sigmoid function to enhance SAR image contrast, and then we separate the target area and remove noise interference. Next, we design a convolutional neural network to obtain the deep features of the target from the separated image. Then, we use four moment methods to produce the moment features of the image and sort them according to recognition rate. Finally, the above features are fused and sorted according to recognition rate to find the optimal combination. The recognition rate of our optimal fusion is 98.63%, which is 2.77% higher than is obtained by CNNs and 6.32% higher than moment methods.