基于DRGAN和支持向量机的合成孔径雷达图像目标识别

为解决SAR图像目标识别中样本缺乏和方位角敏感问题,提出了一种基于DRGAN和SVM的SAR图像目标识别算法。首先,采用多尺度分形特征对SAR图像进行增强,经过分割得到目标二值图像,基于Hu二阶矩估计目标的方位角。然后对估计得到的目标方位角进行量化编码,结合原始图像作为输入,对设计的DRGAN模型参数进行训练与优化。由于DRGAN中的深度生成模型将目标姿态与外观表示进行解耦设计,故可利用该模型将SAR图像样本变换到同一方位角区间。基于变换后的训练样本分别提取归一化灰度特征,利用SVM训练分类器。采用MSTAR数据集在多个不同操作条件下对提出的算法进行测试,实验结果表明,在带变体的标准操作条件下,能够达到97.97%的分类精度,优于部分基于CNN模型的分类精度,在4种扩展操作条件下的分类精度分别为97.83%,91.77%,97.11%和97.04%,均优于传统方法的分类精度。在SAR图像目标方位角估计存在一定误差的情况下,训练得到的GAN模型作为SAR图像目标旋转估计器,能够使得在不进行复杂样本预处理的前提下,仍然取得较高的SAR图像目标识别精度。

SAR ATR based on disentangled representation learning generative adversarial networks and support vector machine

To solve the problem of insufficient training samples and sensitivity of the target’s aspect angle for Synthetic Aperture Radar(SAR) Automatic Target Recognition(ATR), a target recognition algorithm for SAR images based on a DRGAN and Support Vector Machine(SVM) was proposed in this paper. First, a multiscale fractal feature was used to enhance the input SAR image, and the target binary image was obtained through threshold segmentation. Hu second moments were used to estimate the aspect angle of the target. Second, the estimated angle was quantized into a vector, and the parameters of the designed DRGAN model were trained and optimized using these vectors and original images. Because the deep generative model in DRGAN was designed by disentangling the target’s pose from its representation, the aspect angles of SAR image samples could be transformed into the same interval through this model. Normalized gray features were extracted from these transformed training samples, and an SVM classifier was trained using these features. MSTAR database was used to test the performance of the proposed algorithm under different operating conditions. The experimental results demonstrate that the classification accuracy reached 97.97% under standard operating conditions with variants, which is superior to some methods based on a convolutional neural network. The proposed algorithm can achieve classification accuracies of 97.83%,91.77%,97.11% and 97.04% under four extended operating conditions, respectively, which are better than traditional methods. Despite some errors during the estimation of the aspect angle of the object in the SAR image, the trained GAN model acting as rotation estimator of SAR objects still achieves better SAR object recognition performance under the condition in which no complex data preprocessing methods are used.