基于生成对抗网络的图像转换技术

Image transformation technology based on generative adversarial networks

摘要:

针对不同谱段图像获取代价不同的问题，提出一种基于生成对抗网络的图像转换方法。转换过程以肉眼可分辨范围内图像轮廓不变为出发点。首先，通过成对的训练数据对生成器和判别器进行交替训练，不断对损失函数进行优化，直到模型达到纳什平衡。然后用测试数据对上述训练好的模型进行检测，查看转换效果，并从主观观察和客观上计算平均绝对误差和均方误差角度评价转换效果。通过上述过程最终实现不同谱段图像之间的转换。其中，生成器借鉴U-Net架构；判别器采用传统卷积神经网络架构；损失函数方面增加L1损失来保证图像转换前后高、低频特征的完整性。以红外图像与可见光图像之间的转换为例进行实验，结果表明，通过本文设计的生成对抗网络，可以较好地实现红外图像与可见光图像之间的转换。

Abstract:

An image conversion method based on generative adversarial networks is proposed in order to solve the problem of different image acquisition costs in different spectral segments. In the conversion process, the image outline does not change into the starting point in the range that can be distinguished by the naked eye. Firstly, the generator and discriminator are trained alternately through pairs of training data, and the loss function is optimized until the Nash equilibrium of the model is reached. Then the test data are utilized to detect the trained model, to check the conversion effect, and to evaluate the conversion effect from the subjective observation and objective calculation of the average absolute error and mean square error. Through the above process, the conversion between different spectral images is finally realized. Among them, the generator learns from U-Net architecture; the traditional convolution neural network architecture is used by the discriminator; and L1 loss function is increased to ensure the integrity of high and low frequency features before and after image conversion. In this paper, the conversion between infrared image and visible image is taken as an example to carry out the experiment. The results show that the conversion between infrared image and visible image can be well realized through the generative adversarial networks designed in this paper.