【摘要】 针对合成孔径雷达(synthetic aperture radar, SAR)图像样本数据有限,且不同类别间的图像区分度不高导致识别困难的问题,提出一种应用于SAR图像识别的距离度量学习方法.该方法使用CNN网络得到图像的特征分布,利用LSTM网络加强图像间的关联性,基于余弦相似距离度量方法计算图像之间的匹配度,通过注意力机制后对结果进行分类.训练过程结合小样本学习的训练方式,采取预训练的策略进行实验.实验以公开的MSTAR数据集进行SAR图像识别,结果表明该方法准确率达到99.3%,比SVM方法提升2.5%.更多还原

【Abstract】 Due to synthetic aperture radar(SAR)image sample data is insufficient, and the similarity of intra-class images, which causes difficulty in recognition. A distance metric learning method was proposed for SAR image recognition. The method was arranged to use CNN networks to obtain the feature distribution of the image, and use the LSTM networks to strengthen the correlation between images. Based on the cosine similarity distance measurement method, the matching degree between images was calculated, and the results were classified based on the attention mechanism. Combined with the training method of few-shot learning, the training experiments were carried out with pre-training strategies and using the public MSTAR data set to perform SAR image recognition. The results show that the recognition rate of the method can reach up to 99.3%,be 2.5% higher than SVM.更多还原