基于生成对抗网络的图像修复算法

Image Inpainting Algorithm Based on GenerativeAdversarial Network

为了获得更好的图像修复效果,提出了一种基于生成对抗网络的图像修复算法。该算法能在图像像素大量缺失的情况下复原图像,并且更加关注提升修复区域的连续性以及局部一致性。首先使用金字塔上下文编码器对上下文语义进行编码得到潜在特征图;然后通过注意力转移网络学习缺失区域内外的亲和性,再通过加权赋值的方式,将注意力评分转移到潜在特征图中的填充区域,形成重建特征图;接下来将潜在特征图和对应的重建特征图进行跳跃连接,并将其输入到多尺度译码器中生成修复图像;最后通过全局局部一致的鉴别网络来区分真实图像和修复图像。实验结果表明,所提算法在PSNR和SSIM评价指标值均有所提升的同时L1 loss评价指标值有所降低,修复后的图像更接近于真实图像。

In order to obtain better effect of image inpainting,an image inpainting algorithm based on generative adversarial network is proposed.This algorithm can restore the image with a large number of missing pixels and pay more attention to improving the continuity and local consistency of the restored region.Firstly,the pyramid context encoder encodes the context semantics to get the potential feature map.Then,the agreeableness inside and outside the missing area is learned through the attention transfer network.Then,the weighted assignment is used to transfer the attention score to the filled area in the potential feature map to form the reconstructed feature map.Next,the potential feature map and the corresponding reconstructed feature map are jump-connected and input into the multi-scale decoder to generate the restored image.Finally,a globally and locally consistent authentication network is used to distinguish the real image from the restored image.The comparative experimental results show that the proposed algorithm's PSNR and SSIM evaluation indexes are improved.Moreover,the restoration area has better continuity,and the inpainting image is closer to the real image.