Journal of Computer Applications ›› 2020, Vol. 40 ›› Issue (7): 2084-2088.DOI: 10.11772/j.issn.1001-9081.2019122253

• Virtual reality and multimedia computing • Previous Articles     Next Articles

Infrared image data augmentation based on generative adversarial network

CHEN Foji1,2,3,4, ZHU Feng1,2,3,4, WU Qingxiao1,2,3,4, HAO Yingming1,2,3,4, WANG Ende1,2,3,4   

  1. 1. Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang Liaoning 110016, China;
    2. Institutes for Robotics and Intelligent Manufacturing Innovation, Chinese Academy of Sciences, Shenyang Liaoning 110016, China;
    3. University of Chinese Academy of Sciences, Beijing 100049, China;
    4. Key Laboratory of Opto-Electronic Information Processing, Chinese Academy of Sciences, Shenyang Liaoning 110016, China
  • Received:2020-01-09 Revised:2020-03-01 Online:2020-07-10 Published:2020-03-29
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (U1713216).

基于生成对抗网络的红外图像数据增强

陈佛计1,2,3,4, 朱枫1,2,3,4, 吴清潇1,2,3,4, 郝颖明1,2,3,4, 王恩德1,2,3,4   

  1. 1. 中国科学院 沈阳自动化研究所, 沈阳 110016;
    2. 中国科学院 机器人与智能制造创新研究院, 沈阳 110016;
    3. 中国科学院大学, 北京 100049;
    4. 中国科学院 光电信息处理重点实验室, 沈阳 110016
  • 通讯作者: 陈佛计
  • 作者简介:陈佛计(1994-),男,山西忻州人,硕士,主要研究方向:图像生成、机器学习、机器人视觉;朱枫(1962-),男,辽宁沈阳人,研究员,博士生导师,博士,主要研究方向:机器人视觉;吴清潇(1978-),男,辽宁沈阳人,研究员,博士,主要研究方向:机器人视觉、机器视觉;郝颖明(1966-),女,辽宁沈阳人,研究员,博士,主要研究方向:视觉定位、图像优化、视觉检测、红外图像仿真、三维成像、数据处理;王恩德(1980-),男,辽宁沈阳人,研究员,博士,主要研究方向:图像目标识别、检测与跟踪、微弱信号检测。
  • 基金资助:
    国家自然科学基金资助项目(U1713216)。

Abstract: The great performance of deep learning in many visual tasks largely depends on the big data volume and the improvement of computing power. But in many practical projects, it is usually difficult to provide enough data to complete the task. Concerning the problem that the number of infrared images is small and the infrared images are hard to collect, a method to generate infrared images based on color images was proposed to obtain more infrared image data. Firstly, the existing color image and infrared image data were employed to construct the paired datasets. Secondly, the generator and the discriminator of Generative Adversarial Network (GAN) model were formed based on the convolutional neural network and the transposed convolutional neural network. Thirdly, the GAN model was trained based on the paired datasets until the Nash equilibrium between the generator and the discriminator was reached. Finally, the trained generator was used to transform the color image from the color field to the infrared field. The experimental results were evaluated based on quantitative evaluation metrics. The evaluation results show that the proposed method can generate high-quality infrared images. In addition, after the L1 or L2 regularization constraint was added to the loss function, the FID (Fréchet Inception Distance) score was respectively reduced by 23.95, 20.89 on average compared to the FID score of loss function not adding the constraint. As an unsupervised data augmentation method, the method can also be applied to many other visual tasks that lack train data, such as target recognition, target detection and data imbalance.

Key words: infrared image generation, Generative Adversarial Network (GAN), image transformation, data augmentation, quality evaluation of generative image

摘要: 深度学习在视觉任务中的良好表现很大程度上依赖于海量的数据和计算力的提升,但是在很多实际项目中通常难以提供足够的数据来完成任务。针对某些情况下红外图像少且难以获得的问题,提出一种基于彩色图像生成红外图像的方法来获取更多的红外图像数据。首先,用现有的彩色图像和红外图像数据构建成对的数据集;然后,基于卷积神经网络、转置卷积神经网络构建生成对抗网络(GAN)模型的生成器和鉴别器;接着,基于成对的数据集来训练GAN模型,直到生成器和鉴别器之间达到纳什平衡状态;最后,用训练好的生成器将彩色图像从彩色域变换到红外域。基于定量评估标准对实验结果进行了评估,结果表明,所提方法可以生成高质量的红外图像,并且相较于在损失函数中不加正则化项,在损失函数中加入L1和L2正则化约束后,该方法的FID分数值平均分别降低了23.95和20.89。作为一种无监督的数据增强方法,该方法也可以被应用于其他缺少数据的目标识别、目标检测、数据不平衡等视觉任务中。

关键词: 红外图像生成, 生成对抗网络, 图像转换, 数据增强, 生成图像质量评估

CLC Number: