基于新生成器结构的图像修复方法

针对目前图像修复算法存在的修复效果不连续、缺失大小受限、训练过程不稳定等缺点,提出了一种基于生成对抗网络的图像修复方法.利用卷积神经网络,我们可以真实地修复任意分辨率的图像.为了实现高分辨率的真实修复效果和对图像特征的充分学习,我们提出基于DenseNet传播源图像的细节和结构得到高分辨率的图像,实现图像缺失生成;由于Iizuka等人提出的基于双判别器方法中膨胀卷积部分所产生的巨大运算量,我们提出使用JPU (Joint Pyramid Upsampling,联合金字塔上采样)来加速计算.在CelebA和ImageNet上的实验表明,所提方法能真实地修复大多数的破损图像.

Image Restoration Method Based on New Generator Structure

Aiming at the shortcomings of current image restoration algorithms, such as discontinuity of repair effect, limitation of missing size, and instability of training process, an image restoration method based on generation antagonistic network is proposed. Using convolutional neural network, we can actually repair images of any resolution. In order to realize the real restoration effect of high resolution and the full learning of image features, we propose to obtain high resolution images based on the details and structure of DenseNet propagating source images, so as to realize the generation of missing images. As Iizuka et al. proposed the large computation amount generated by the expanded convolution part in the two-discriminator method, we propose to use JPU (Joint Pyramid Upsampling) to accelerate the calculation. Experiments in CelebA and ImageNet show that the proposed method can truly repair most broken images.