图像修复的应用领域非常广泛，在图像的传播扩散或者是保存的过程中，常有损坏、失真或者受到污染等问题，从而导致图像信息不完整，所以，在图像处理技术当中，图像修复技术属于一项十分重要的研究工作。传统的图像修复技术是基于破损区域的周围的低级图像特征，在已知区域搜索相似内容进行修复，不能理解高级语义，很容易造成视觉上的偏差。随着深度学习技术的不断完善和发展，通过卷积神经网络技术学习了解图像的特征，能够更好地修复图像中破损的部分，相对于传统的修复方法，在图像修复技术上做了改进，但其修复结果仍然存在着伪影、具体细节模糊、失真等问题。

为解决这些问题，本文从深度学习原理和现有的修复模型出发，以生成对抗网络为基础，将其进行改进，改进出一种新的基于生成对抗网络的图像修复算法，主要的研究内容有以下几个方面：

第一，采用双阶段生成器网络，第一阶段生成器为结构重建网络，引入残差思想并加入自注意力机制，通过跳跃层不断强调初始颜色信息，最大程度加强色差；第二阶段生成器为颜色校正网络，将第一阶段的结果结合掩码图作为输入，进行颜色校正。两个生成器共享一个谱归一化马尔可夫判别器。生成器第二阶段为颜色校正网络，所以采用直方图损失，减少生成图像与真实图像的色差。生成器损失包括感知损失、L1损失以及直方图损失三个损失函数。

第二，采用残差门控卷积替代原先结构重建网络中残差网络的普通卷积层，门控卷积对有效和无效像素特殊的处理方法，在部分像素缺失和大尺度损坏情况下，能够更好地对局部特征进行精确描述。在颜色校正网络中引入通道注意力模块提取色彩信息，通道注意力模块能够根据图像特征的不同赋予不同的权重，对关联性强的特征赋予较高权重，增强其辅助作用，对关联性弱的特征赋予低权重，抑制其对缺失区域内容的影响。

第三，为了验证本论文模型的准确性，结合使用主观、客观的评价标准，对本模型的图像修复结果进行评价。从主观和客观两个角度对比不同算法在破损区域不同比例下的不同图像的修复结果，相对于之前的两阶段修复，本文所改进的算法在峰值信噪比上提升了8.1%，在结构相似性上提升了4.0%，并且主观上视觉效果明显。

The application field of image inpainting is very wide. In the process of image propagation, diffusion or preservation, problems such as damage, distortion or pollution often occur, which lead to incomplete image information. Therefore, among image processing technologies, inpainting technology is an important research work. The traditional inpainting technology is based on the low-level image features around the damaged area, searching for similar content in the known area to repair, unable to understand the high-level semantics, which is easy to cause visual deviation. With the continuous improvement and development of deep learning technology, the convolutional neural network technology can learn and understand the features of images to better inpainting the damaged parts of images. Compared to traditional inpainting methods, the inpainting method based on deep learning has made improvements in inpainting technology, but its repair results still have artifacts, blurred details, distortion and other problems.

To solve these issues, this thesis starts from the principles of deep learning and existing image inpainting models, based on generative adversarial networks, improves them, and improves a new image repair algorithm based on generative adversarial networks. The main research content includes the following aspects:

Firstly, this thesis chooses to use a two-stage generator network, the first stage of the generator is a structural reconstruction network, introducing residual thinking and incorporating a self-attention mechanism. By continuously emphasizing initial color information through skip layers, the color difference is maximized; The second stage generator is a color correction network, which combines the results of the first stage with a mask map as input to perform color correction. Two generators share a spectral normalized patch GAN discriminator. The second stage of the generator is a color correction network, which uses histogram loss to reduce the color difference between the generated image and the real image. The overall loss function consists of perceptual loss, L1 loss and histogram loss.

Secondly, this thesis uses residual gated convolution to replace the ordinary convolutional layer of the residual network in the original structural reconstruction network, the unique effective pixel and invalid pixel processing methods of the gated convolution can better solve the local feature accurate description in the case of partial pixel loss and large-scale damage in the application of inpainting. Introducing channel attention module to extract color information in color correction networks, the channel attention module can assign different weights based on different image features, assign higher weights to features with strong correlation, enhance their auxiliary effect, assign low weights to features with weak correlation, and suppress their impact on missing area content.

Thirdly, in order to verify the accuracy of the model in this thesis, a combination of subjective and objective evaluation criteria will be used to evaluate the image restoration results of this model. By comparing the repair results of different images in different proportions of missing areas with other algorithms, and verifying them from both subjective and objective perspectives, it is fully proved that the proposed algorithm has good inpainting effect.

compared to the previous two-stage image inpainting model, the improved algorithm in this thesis has increased the peak signal to noise ratio by 8.1% and the structural similarity index by 4.0%, and the visual effect is obvious subjectively.