Scholarlink

<p>Shallow relief is one of the oldest art forms in human civilization, and its effect generation is beneficial for the transmission of cultural genes. To improve the generation effect of shallow relief, an improved U-shaped network is designed for normal estimation to accurately extract key information from the input image. Subsequently, the study employs a hybrid dilation and attention residual U-shaped network, incorporating constraints on normal consistency loss and structural similarity loss to generate shallow relief of geometric shapes. A normal style transfer method based on Swin Transformer is proposed and integrated with the aforementioned model to form the final generation technique. The results show that the average accuracy of normal information generation in the improved U-shaped network is 92.3%, which is much higher than other module combinations, and its maximum time consumption is 20.3&#xa0;ms, with little difference compared to other module combinations. The geometric form of shallow relief is closer to the real target, with an average relief depth error of 0.316&#xa0;mm, far lower than the comparative models of 0.850&#xa0;mm, 0.776&#xa0;mm, 0.646&#xa0;mm, and 0.576&#xa0;mm. In addition, the average content retention of normal style transfer is 91.32%, and the average geometric error in practical application is 0.18&#xa0;mm. The shallow relief generation technology has good performance and can help inherit and innovate cultural genes.</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

Shallow relief effect generation technology integrating improved U-Net and normal style transfer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月23日

查看全文
  • Xingshan Liu

摘要

Shallow relief is one of the oldest art forms in human civilization, and its effect generation is beneficial for the transmission of cultural genes. To improve the generation effect of shallow relief, an improved U-shaped network is designed for normal estimation to accurately extract key information from the input image. Subsequently, the study employs a hybrid dilation and attention residual U-shaped network, incorporating constraints on normal consistency loss and structural similarity loss to generate shallow relief of geometric shapes. A normal style transfer method based on Swin Transformer is proposed and integrated with the aforementioned model to form the final generation technique. The results show that the average accuracy of normal information generation in the improved U-shaped network is 92.3%, which is much higher than other module combinations, and its maximum time consumption is 20.3 ms, with little difference compared to other module combinations. The geometric form of shallow relief is closer to the real target, with an average relief depth error of 0.316 mm, far lower than the comparative models of 0.850 mm, 0.776 mm, 0.646 mm, and 0.576 mm. In addition, the average content retention of normal style transfer is 91.32%, and the average geometric error in practical application is 0.18 mm. The shallow relief generation technology has good performance and can help inherit and innovate cultural genes.