基于密集卷积生成对抗网络的图像修复

Dense Convolution Generative Adversarial Networks Based Image Inpainting

图像修复是一项利用缺损图像中已知信息对缺损区域信息进行估计修复的技术。针对大面积语义信息缺失的图像进行修复时,若训练数据集较小且图像背景相对复杂,则基于生成模型的修复结果常出现模糊、伪影和视觉相似度差等问题。针对上述问题,文中提出了一种基于密集卷积生成对抗网络的图像修复算法。该算法采用生成对抗网络作为图像修复的基本框架。首先,利用密集卷积块构建具有编解码结构的生成网络,不但加强了图像特征的提取,提高了图像修复能力,而且避免了深度增加引起的梯度消失问题。其次,在编码和解码结构之间引入跳跃连接,解决了网络层间信息传递丢失的问题。然后,在网络优化过程中,结合重建损失、对抗损失和TV损失来训练网络模型,增强了网络稳定性。最后,分别在CelebA和Car两个数据集上进行实验,所提算法的修复结果在视觉效果、峰值信噪比PSNR和结构相似度SSIM 3个方面均优于3种代表性图像修复算法,其有效性得到验证。

Image inpainting is one technique of reconstruction defect areas by inferring information from the known context of defect images.For semantic image inpainting of large areas,there are still many problems in inpainting algorithms based on generation models,such as blur,artifacts,and poor visual similarity,especially for the complex background images and small datasets.To solve this problem,an image inpainting algorithm based on dense convolution generative adversarial networks is proposed.The generated adversarial network is the basic framework.Firstly,dense convolutional blocks are used to enhance image feature extraction,improve image repair capability,and avoid the problem of gradient disappearance caused by the network depth increa-sing in the generator network.Secondly,skip connection between the encoding and decoding structures is involved to avoid information transmission lost problems between network layers.After that,a total loss function,composed of the reconstruction loss,adversarial loss and TV loss,is used to optimize the network and enhance network stability.Finally,the proposed algorithm is validated on the CelebA dataset and Car dataset respectively,compared with three typical image inpainting algorithms.The effectiveness of the algorithm is proved in visual perception,peak signal-to-noise ratio(PSNR)and structural similarity(SSIM).