Optimal azimuth angle selection for limited SAR vehicle target recognition

Lack of labeled data is a common problem among synthetic aperture radar (SAR) target recognition, which can be defined as few-shot and limited-data SAR target recognition. The low accuracy under scarce labeled data is mainly due to the sensitivity of azimuth angles from the labeled data. When the labeled data is abundant and covers the most angles, the recognition rate can easily reach a high level (over 90%) no matter what recognition algorithm it is. However, in actual missions, SAR target recognition usually cannot achieve enough labeled data and is a few-shot or limited-data recognition problem in most cases. Thus, under these circumstances, the problem of samples with certain azimuth angles being selected to maximize recognition rate with fixed number labeled data is studied in this article. Based on this point, the basic optimization model is established, and the concept of boundary performance is proposed. Furthermore, the goal of the proposed method is to select the most representative N-shot samples based on the azimuth angle to make the final recognition rate as high as possible. The motivation of the proposed representative azimuth angle selection algorithm is to select the samples, whose feature similarities change gently to candidate more samples in embedding space. Through abundant contrast experiments, the representative samples selected from the proposed method achieve the highest recognition rate. According to the contrast experiments, the samples selected by our proposed method exceed random selection manner more than 7% and 10 % in PD Network and A-ConvNet, respectively, in 10-way 10-shot. In addition, through this article, the boundary performance among SAR target recognition can be defined, and that is to say, how many samples are recognized as a few-shot recognition problem in SAR target classification is ensured in this article.