基于生成对抗网络的遥感影像色彩一致性方法

A method for color consistency of remote sensing images based on generative adversarial networks

在遥感成像过程中易在拍摄影像内部、影像与影像之间产生亮度不均匀、色彩不一致的现象,通过人工借助图像处理软件进行色彩调节已经不能满足呈几何级数量增长的遥感影像调色需求,因此提出一种针对土地利用率高的复杂城区地物的融合注意力机制无监督循环一致生成对抗网络(channel attention-cycle generative adversarial networks,CA-CycleGAN)。首先,通过直方图调整和Photoshop等软件手工制作用于色彩参考的样本数据集,选择合适的城区影像数据作为待校正影像样本集,将2部分影像分别进行裁切,得到预处理后的影像样本集;然后,将处理好的待校正影像集和色彩参考影像集通过CA-CycleGAN中,由于在生成器中加入了注意力机制,因此在生成器与鉴别器相互对抗的训练过程中能够利用注意力特征图将生成的重点分配在重要的区域,提高生成影像效果,得到基于城区影像的色彩校正模型以及色彩校正后的影像图。影像校正效果和损失函数图表明,所提出的方法在循环一致生成对抗网络基础上做出了优化,加入注意力机制的生成对抗网络在调整影像色彩上的综合表现效果优于不加注意力机制的生成对抗网络。相较于传统方法大大减少了色彩校正的时间,对比人工调色增加了影像色彩校正效果的稳定性。证明所提出方法在遥感影像匀色工作中优势较明显,具有较好的应用前景。

Uneven brightness and inconsistent colors are prone to occur inside and between captured images in the process of remote sensing imaging.However,the manual color conditioning combined with image processing software can no longer meet the color matching demand of geometrically increasing remote sensing images.Given this,this study proposed a kind of unsupervised channel-cycle generative adversarial network(CA-CycleGAN)integrated with the attention mechanism suitable for ground objects in complex urban areas with a high land utilization rate.Firstly,the sample data set used for color reference was manually made through histogram adjustment and Photoshop,and the appropriate urban images were selected as the sample set to be corrected.Then,the two kinds of images were cut respectively to obtain the preprocessed image sample sets.Finally,the preprocessed image set to be corrected and the image set for color reference were processed using the CA-CycleGAN.Because the attention mechanism has been added to the generator,the generated focuses can be distributed into key areas using the attention feature map in the training process of the confrontation between the generator and the discriminator,thus improving the image effects and obtaining the color correction model based on urban images and the images after color correction.Both the image correction effect and the loss function diagram show that the proposed method is optimized based on the CycleGAN and that the comprehensive performance of the CycleGAN integrated with the attention mechanism is better than that without the attention mechanism.Compared to conventional methods,the method proposed in this study greatly reduced the time for color correction and achieved more stable image color correction effects than manual color matching.Therefore,the method proposed in this study enjoys significant advantages in the color dodging of remote sensing images and has a good application prospect.