基于卷积对抗网络的多通道图像修复方法

MULTI-CHANNEL IMAGE RESTORATION METHOD BASED ON CONVOLUTIONAL ADVERSARIAL NETWORKS

基于对抗网络与图像多通道卷积的思想,提出一种图像修复算法。卷积层通过分离通道的方式对子通道卷积编码,单通道以多层次编码的结构融入了浅层表示以及不同特征图的信息。设计了合理的网络架构,分析了损失函数的优化,通过生成对抗网络损失、图像级重建损失和通道损失组合进行训练。实验结果表明,该算法可以得到视觉上更加自然的重建图像,峰值信噪比与结构相似性的指标相比于现阶段的修复算法也有可观的提升。

Based on the idea of adversarial networks and the multi-channel convolution of images,we propose an image restoration algorithm.The convolutional layer encoded the sub-channel by separating channels,and the single channel integrated the information of shallow representation and different feature graph in a multi-level coding structure.We designed a reasonable network architecture and analyzed the optimization of loss function.Training was performed by generating a combination of adversarial network loss,image-level reconstruction loss and channel loss.The experimental results show that our algorithm can achieve better visual and natural reconstruction images,and the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)index are significantly improved compared with the current restoration algorithm.