摘要
针对航空图像中厚云去除的难题,提出一种基于深度卷积生成对抗网络的航拍图像去厚云方法。将图像中被云遮挡的区域看作图像修复问题中的缺失部分,利用卷积神经网络的对抗学习补偿缺失信息。设计了包括生成器-鉴别器的深度卷积生成对抗网络模型。生成器采用编码器-解码器结构,构建了包含重建损失、对抗损失和总变差损失的联合损失函数,不断训练以生成云区的预测图像;鉴别器衡量生成图像的真实性,以对抗损失作为损失函数。通过不断迭代联合优化生成器和鉴别器,以使网络预测性能提高。引入泊松图像编辑平滑边界,以降低颜色差异和边界伪迹的影响。在模拟含云图像与真实含云图像上实验结果表明,所提出方法的去云效果在峰值信噪比、结构相似性、自然图像无参考质量评价算法及其改进算法指标优于经典方法,更符合人眼主观感受,且具有较小的运算复杂度。
A thick cloud removal method for aerial images based on deep convolution generative adversa- rial network (DCGAN) is proposed. In the proposed method, a region covered by cloud in image is regarded as a missing part in the image inpainting, and the adversarial learning of convolutional neural network is used to compensate the missing information. A DCGAN model including generator- discriminator is designed. The generator is an encoder-decoder structure with a joint loss function, including reconstruction loss, adversarial loss and total variation loss. The discriminator is used to measure the authenticity of the generated image with a loss function of adversarial loss. The prediction performance of network is improved by iterating the joint optimization of generator and discriminator. Poisson image editing is introduced to reduce the influences of color difference and boundary artifact. The experimental results on both simulated and real cloud images show that the proposed method is better than the classic method in terms of peak signal to noise ratio, structural similarity index measure, natural image no-reference quality evaluation algorithm and its improved algorithm, and has a small computational complexity.
作者
李从利
张思雨
韦哲
薛松
LI Congli;ZHANG Siyu;WEI Zhe;XUE Song(PLA Army Academy of Artillery and Air Defense, Hefei 230031, Anhui, China)
出处
《兵工学报》
EI
CAS
CSCD
北大核心
2019年第7期1434-1442,共9页
Acta Armamentarii
关键词
航拍图像
厚云去除
深度卷积生成对抗网络
泊松图像编辑
aerial image
thick cloud removal
deep convolutional generative adversarial network
Poisson image editing