Scholarlink

<p>To address the issue of image quality degradation caused by glass cover reflections during the digitization of Thangka paintings, this paper introduces an innovative iterative prediction network (MPGINet) guided by reflection perception and based on multi-dimensional deep features. The proposed network enhances the robustness of reflection removal by leveraging deep multi-dimensional features from the initially predicted reflection layer, transmission layer, and the original input image. This is further augmented by the strategic integration of frequency-domain information separation and mask-guided image inpainting. The proposed model employs a two-stage iterative architecture. In the first stage, it integrates the U-Net framework with the Squeeze-and-Excitation module, enhanced by a residual mechanism, to refine the prediction of the reflection layer. In the second stage, the network utilizes the novel Deep Feature Pyramid Network (DFPN), which excels in capturing the fine-grained texture features of Thangka paintings. The DFPN effectively fuses high- and low-frequency features under mask constraints, enabling precise restoration of the transmission layer. This dual-stage approach ensures a comprehensive and detailed recovery of the original artwork’s visual fidelity. Experimental results show that on the Thangka datasets, the PSNR and SSIM of MPGINet reach 28.90&#xa0;dB and 0.962 respectively, an increase of 1.88&#xa0;dB and 0.027 compared with the existing state-of-the-art (SOTA) methods. In natural scene datasets, the proposed method achieves comparable results with SOTA methods, verifying its generalization ability.</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

Multi-dimensional perception-guided iterative reflection removal network with deep features for painting images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月29日

查看全文
  • Yuqi Xie,
  • Xiaojuan Zhang,
  • Yang Zhao

摘要

To address the issue of image quality degradation caused by glass cover reflections during the digitization of Thangka paintings, this paper introduces an innovative iterative prediction network (MPGINet) guided by reflection perception and based on multi-dimensional deep features. The proposed network enhances the robustness of reflection removal by leveraging deep multi-dimensional features from the initially predicted reflection layer, transmission layer, and the original input image. This is further augmented by the strategic integration of frequency-domain information separation and mask-guided image inpainting. The proposed model employs a two-stage iterative architecture. In the first stage, it integrates the U-Net framework with the Squeeze-and-Excitation module, enhanced by a residual mechanism, to refine the prediction of the reflection layer. In the second stage, the network utilizes the novel Deep Feature Pyramid Network (DFPN), which excels in capturing the fine-grained texture features of Thangka paintings. The DFPN effectively fuses high- and low-frequency features under mask constraints, enabling precise restoration of the transmission layer. This dual-stage approach ensures a comprehensive and detailed recovery of the original artwork’s visual fidelity. Experimental results show that on the Thangka datasets, the PSNR and SSIM of MPGINet reach 28.90 dB and 0.962 respectively, an increase of 1.88 dB and 0.027 compared with the existing state-of-the-art (SOTA) methods. In natural scene datasets, the proposed method achieves comparable results with SOTA methods, verifying its generalization ability.