小样本条件下SCGAN+CNN低分辨雷达目标一步识别算法

现有低分辨雷达目标识别方法，通常采用先特征提取、再进行目标分类的两步识别算法，这种算法存在识别率难以提高和方法泛化性不足的问题，对此，提出一种增强条件生成对抗网络（strengthening condition generative adversarial network，SCGAN）+卷积神经网络（convolutional neural network，CNN）的低分辨雷达目标一步识别算法。该算法利用CNN自动获取采样数据深层本质特征，无需特征提取，实现对目标的一步识别。为进一步提高小样本条件下的识别效果，基于CGAN理论来提高样本在特征空间的覆盖程度，并对CGAN的判别器进行改进，在损失函数中增加混叠惩戒项，通过SCGAN生成不混叠的生成样本来更好地训练CNN，提高其在小样本条件下的识别能力。仿真对比实验校验了一步识别算法较传统两步识别算法的优越性，以及SCGAN+CNN的低分辨雷达目标一步识别算法在小样本条件下的有效性。

Low-resolution radar target one-step recognition algorithm based on SCGAN+CNN with a limited training data

The traditional low-resolution radar target recognition technology has the difficulty in improving the recognition rate and insufficient generalization because of using a two-step recognition algorithm based on feature extraction and target classification. A low-resolution radar target one-step recognition algorithm based on strengthening condition generative adversarial network (SCGAN)+convolution neural network (CNN) is put forward. CNN is used in the algorithm to automatically obtain the sampling data deep essence characteristics to achieve the low-resolution radar target one-step recognition. In order to improve the recognition rate with limited training data, CGAN theory is used to improve the coverage of sampling in feature space. Moreover, an SCGAN model is proposed to generate samples with unmixed distribution by adding mixed punishment to the discriminator. CNN can better identify radar targets based on SCGAN. The numerical simulations have demonstrated the effectiveness of the low-resolution radar target one-step recognition algorithm and the recognition algorithm based SCGAN+CNN.