Scholarlink

Using Generative Adversarial Networks (GANs) techniques, significant breakthroughs have been achieved in current cutting-edge image restoration schemes for arbitrary missing regions. However, the quandary about generating unrealistic structures and hazy texture effects in high-resolution images still exists due to the lack of reasoning about the remote contextual information of the image and the loss of image details due to the convolution operation. To address the above problem, a GAN based on U-Net using improved aggregated context transformation (UAT-GAN) is proposed. An encoder-decoder in the form of U-Net was designed to build the middle layer of the generator of UAT-GAN with AOT-Plus blocks. This network improves the quality of image completion and makes the results more realistic and natural. Evaluating the model on the datasets CelebHQ, Places2, and QMUL-OpenLogo, quantitative and qualitative results demonstrate that our model can compete with advanced models in the vast majority of aspects.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improved Aggregated Contextual Transformations Based on U-Net for Image Inpainting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shufan Li,
  • Fang Yang

摘要

Using Generative Adversarial Networks (GANs) techniques, significant breakthroughs have been achieved in current cutting-edge image restoration schemes for arbitrary missing regions. However, the quandary about generating unrealistic structures and hazy texture effects in high-resolution images still exists due to the lack of reasoning about the remote contextual information of the image and the loss of image details due to the convolution operation. To address the above problem, a GAN based on U-Net using improved aggregated context transformation (UAT-GAN) is proposed. An encoder-decoder in the form of U-Net was designed to build the middle layer of the generator of UAT-GAN with AOT-Plus blocks. This network improves the quality of image completion and makes the results more realistic and natural. Evaluating the model on the datasets CelebHQ, Places2, and QMUL-OpenLogo, quantitative and qualitative results demonstrate that our model can compete with advanced models in the vast majority of aspects.