基于生成对抗网络的图像修复算法

为解决当前基于生成对抗网络的深度学习网络模型在面对较复杂的特征时存在伪影、纹理细节退化等现象, 造成视觉上的欠缺问题, 提出了连贯语义注意力机制与生成对抗网络相结合的图像修复改进算法. 首先, 生成器使用两阶段修复方法, 用门控卷积替代生成对抗网络的普通卷积, 引入残差块解决梯度消失问题, 同时引入连贯语义注意力机制提升生成器对图像中重要信息和结构的关注度; 其次, 判别器使用马尔可夫判别器, 强化网络的判别效果, 将生成器输出结果进行反卷积操作得到最终修复后的图片. 通过修复结果以及图像质量评价指标与基线算法进行对比, 实验结果表明, 该算法对缺失部分进行了更好地预测, 修复效果有了更好的提升.

Image Inpainting Algorithm Based on Generative Adversarial Network

The current depth learning network models based on generative adversarial networks encounter artifacts, texture detail degradation, and other phenomena when facing more complex features, which leads to a visual deficiency. In order to solve these problems, an improved image inpainting algorithm combining a generative adversarial network with a coherent semantic attention mechanism is proposed. First of all, the generator adopts a two-stage inpainting method, uses gated convolution to replace the ordinary convolution of the generative adversarial network, and introduces residual blocks to solve the gradient vanishing problem and a coherent semantic attention mechanism to enhance the generator''s attention to the important information and structure in the image. Secondly, Markov discriminator is adopted to enhance the network''s discrimination effect, and the output results of the generator are processed by deconvolution to get the final repaired image. By comparing the inpainting results and image quality evaluation indicators with the baseline algorithm, the experimental results show that the algorithm can better predict the missing parts and improve the inpainting effect.