提出一种基于深度学习的红外目标建模方法。将对抗与自编码相结合,设计了双重对抗自编码网络。利用训练后的网络,仅需输入类别标签和满足一定分布的随机变量即可生成相应类别的红外目标图像。在自建红外数据集上对模型的有效性进行验证,实验表明,生成的目标图像在真实性和多样性等各方面均取得了较高的评价结果。将随机生成的目标图像作为小数据集的补充,可有效改善训练数据匮乏的问题,提高红外成像系统识别算法的准确率。

In this study, we propose an infrared target modeling method based on deep learning. Further, we design a conditional double adversarial autoencoding network by combining the adversarial concepts with autoencoding. By using the trained network, the expected infrared target images can be easily generated by inputting category labels and the random variables that satisfy a certain distribution. The effectiveness of the proposed model is verified using a self-built infrared dataset. The conducted experiments prove that the generated target images exhibit considerable authenticity and diversity. Finally, the randomly generated target images as supplement the small data set can effectively improve the problem of lack of training data and improve the accuracy of the recognition algorithm in the infrared imaging system.