Scholarlink

<p>Aiming at problems such as the poor effect of large-hole image inpainting, the limitation of local information reconstruction of the convolutional neural network, and a surge in parameters caused by stacking a large number of convolutional modules in the network model. We make full use of the advantages of convolutional neural network and transformer and propose an improved multi-stage inpainting method with hybrid CNN and transformer. The method achieves a balance between performance and parameters. Specifically, rough results are first generated using a coarse inpainting network with skip connections and lightweight Taylor transformer modules. Then, a local refinement network with coordinate attention is used to perform local refinement, optimizing local details while weakening the influence of unreasonable content in the distance. Finally, to compensate for the inability of local refinement networks to refine the overall pattern over long distances, global refinement is performed using a network that is consistent with the structure of the coarse inpainting network to make the reconstructed image more realistic and natural. Results show that the method outperforms the state of the arts on three publicly available datasets. The code is made available at <a href="https://github.com/Sheeran2000/IMIHCT">https://github.com/Sheeran2000/IMIHCT</a>.</p>

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

IMIHCT: improved multi-stage image inpainting with hybrid CNN and transformer

  • 期刊论文

  • 发表日期：

    2025年1月27日

查看全文
  • Tao Ning,
  • Xingfang Wang,
  • Hongwei Ding

摘要

Aiming at problems such as the poor effect of large-hole image inpainting, the limitation of local information reconstruction of the convolutional neural network, and a surge in parameters caused by stacking a large number of convolutional modules in the network model. We make full use of the advantages of convolutional neural network and transformer and propose an improved multi-stage inpainting method with hybrid CNN and transformer. The method achieves a balance between performance and parameters. Specifically, rough results are first generated using a coarse inpainting network with skip connections and lightweight Taylor transformer modules. Then, a local refinement network with coordinate attention is used to perform local refinement, optimizing local details while weakening the influence of unreasonable content in the distance. Finally, to compensate for the inability of local refinement networks to refine the overall pattern over long distances, global refinement is performed using a network that is consistent with the structure of the coarse inpainting network to make the reconstructed image more realistic and natural. Results show that the method outperforms the state of the arts on three publicly available datasets. The code is made available at https://github.com/Sheeran2000/IMIHCT.