Attribute Scattering Center-Assisted SAR ATR Based on GNN-FiLM
Junjie Hou,Zhidan Bian,Gengjiang Yao,Hai Lin,Yunhua Zhang,Siyuan He,Haitao Chen
DOI: https://doi.org/10.1109/lgrs.2024.3391323
IF: 5.343
2024-05-10
IEEE Geoscience and Remote Sensing Letters
Abstract:Machine learning-assisted synthetic aperture radar (SAR) automatic target recognition (ATR) methods have gained widespread attention and application. Among them, the methods based on the convolutional neural network (CNN) have achieved good results under standard operating conditions (SOCs) because the CNN can automatically extract and learn the image feature from the SAR image. Still, they show poor performance under extended operating conditions (EOCs) due to the change of the radar imaging conditions, especially when the depression angle of radar imaging changes a lot. Compared to the image features, the scattering center (SC) can provide a concise and physically relevant description of the target scattering characteristics. To achieve a good recognition performance under both SOC and EOC, we propose an SAR ATR method based on the SC features and graph neural networks with featurewise linear modulation (GNN-FiLM). By quantifying and representing scattering features with the physical parameters of SCs, the SC features are introduced into the training of GNN-FiLM in the form of a graph. The structure features and physical information are extracted and learned by the GNN-FiLM. Experimental results on the Moving and Stationary Target Acquisition and Recognition (MSTAR) and simulation datasets demonstrate that our method can perform well in recognition experiments of three-class targets. Especially under EOC of large depression variant on the MSTAR dataset, our method can obtain about 91.5% accuracy, which is better than other deep learning-based methods.
imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geochemistry & geophysics