光电信息技术研究室知识库

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。该方法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生成对抗网络模型，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。实验结果表明，该方法可以生成高质量的红外图像，并且基于定量的评估方法（FID）对实验结果进行了验证，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，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。

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