周立君, 刘宇, 白璐, 茹志兵, 于帅. 一种基于GAN和自适应迁移学习的样本生成方法[J]. 应用光学, 2020, 41(1): 120-126. DOI: 10.5768/JAO202041.0102009
引用本文: 周立君, 刘宇, 白璐, 茹志兵, 于帅. 一种基于GAN和自适应迁移学习的样本生成方法[J]. 应用光学, 2020, 41(1): 120-126. DOI: 10.5768/JAO202041.0102009
ZHOU Lijun, LIU Yu, BAI Lu, RU Zhibing, YU Shuai. Sample generation method based on GAN and adaptive transfer learning[J]. Journal of Applied Optics, 2020, 41(1): 120-126. DOI: 10.5768/JAO202041.0102009
Citation: ZHOU Lijun, LIU Yu, BAI Lu, RU Zhibing, YU Shuai. Sample generation method based on GAN and adaptive transfer learning[J]. Journal of Applied Optics, 2020, 41(1): 120-126. DOI: 10.5768/JAO202041.0102009

一种基于GAN和自适应迁移学习的样本生成方法

Sample generation method based on GAN and adaptive transfer learning

  • 摘要: 研究了基于生成式对抗网络(GAN)和跨域自适应迁移学习的样本生成和自动标注方法。该方法利用自适应迁移学习网络,基于已有的少量可见光图像样本集,挖掘目标在红外和可见光图像中特征内在相关性,构建自适应的转换迁移学习网络模型,生成标注好的目标图像。提出的方法解决了红外图像样本数量少且标注费时的问题,为后续多频段协同目标检测和识别获得了足够的样本数据。实验结果表明:自动标注算法对实际采集的装甲目标图像和生成的装甲目标图像各1 000张进行自动标注测试,对实际装甲目标图像的标注准确率达到95%以上,对生成的装甲目标标注准确率达到83%以上;利用真实图像和生成图像的混合数据集训练的分类器的性能和使用纯真实图像时基本一致。

     

    Abstract: The method of sample generation and automatic annotation based on the generative countermeasure network (GAN) and cross-domain adaptive transfer learning was studied. In this method, the adaptive transfer learning network is used to explore the intrinsic correlation of target features in infrared and visible images based on the small number of existing visible image samples, and the adaptive transfer learning network model is constructed to generate tagged target images. The problem of small number of infrared image samples and time-consuming labeling can be solved by proposed method, which provides enough sample data for subsequent multi-band cooperative target detection and recognition. Moreover, automatic standard tests were carries out on the 1 000 pieces of actual acquired and 1 000 pieces of generated armored target images ,respectively,by using the automatic standard algorithm The experimental results show that the accuracy of the actual armored target image labeling is more than 95%, and that of the generated armored target image labeling is more than 83%. The performance of classifiers trained with the mixed dataset of real images and generate images is basically the same as when using the pure real images.

     

/

返回文章
返回