| 引用本文: | 李筱玉,张 乾,徐开丽,骆 迪,冉娅琴.基于双判别器的 Wasserstein 生成对抗网络图像修复(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2025，42（4）：9-16 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 目的 针对基于生成对抗网络的图像修复模型,在修复不规则大面积缺损图像时,存在细节特征信息还原不
准确、图像视觉效果欠佳及训练不稳定等缺陷问题,提出了一种基于双判别器的 Wasserstein 生成对抗网络图像修
复模型。 方法 该方法首先以编码器-解码器结构的卷积神经网络作为生成器,在生成器的编码器和解码器之间添
加跳跃连接,以更好地学习图像细微特征并提升最终的修复效果;接着在全局判别器基础上引入局部判别网络,以
保证局部修复结果与周围区域的一致性,并在判别器中引入 Wasserstein 距离,从而使网络的训练更加稳定以获得
较为自然的图像修复效果;最后设计 VGG16 特征提取器以引入感知损失和风格损失,通过添加多个损失函数来提
升图像复原效果。 结果 在公开人脸、场景和街景数据集上进行对比分析,验证出所提出的方法定性和定量分析结
果均优于对比模型,其性能评价指标值更高。 结论 实验结果表明:所提出的方法在修复不规则大面积缺损区域的
图像时,修复图像更清新,视觉效果更好。 |
| 关键词: 图像修复 生成对抗网络 跳跃连接 VGG-16 特征提取模型 Wasserstein 距离 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Image Inpainting Using Wasserstein Generative Adversarial Networks with Dual Discriminators |
|
LI Xiaoyu ZHANG Qian XU Kaili LUO Di RAN Yaqin
|
|
School of Data Science and Information Engineering Guizhou Minzu University Guiyang 550025 China
|
| Abstract: |
| Objective Image inpainting models based on generative adversarial networks GANs suffer from deficiencies
when repairing irregular and extensively damaged images including inaccurate restoration of fine image details
unsatisfactory visual effects and training instability. In this regard a Wasserstein GAN-based image inpainting model with
dual discriminators was proposed. Methods This method employed a convolutional neural network CNN with an encoder-
decoder architecture as the generator with skip connections added between the encoder and decoder of the generator to
better learn subtle image features and enhance the final inpainting results. Additionally a local discriminator network was
introduced on the basis of the global discriminator to ensure consistency between locally restored areas and their
surroundings. Wasserstein distance was incorporated into the discriminators to stabilize the training process and achieve more
natural image inpainting. Furthermore a VGG16 feature extractor was designed to introduce perceptual and style losses to
improve image inpainting by incorporating multiple loss functions. Results Comparative analysis on public datasets of
faces scenes and streets showed qualitative and quantitative superiority of the proposed method over baseline models with
higher performance evaluation metrics. Conclusion Experimental results demonstrate that the proposed method presents
better visual effects and clearer restoration when repairing images with irregular and extensive damaged areas. |
| Key words: image inpainting generative adversarial networks skip connections VGG-16 feature extractor Wasserstein
distance |
