Scholarlink

<p>The maximum intensity and gradient values of non-overlapping patches significantly decrease during the blurring process. To address this issue, we propose an enhanced patch-wise maximum gradient (<i>EPMG</i>) prior for blind image deblurring. We evaluate the statistical distribution of <i>EPMG</i> using a real dataset and mathematically demonstrate its effectiveness. Based on the <i>EPMG</i> prior, we develop an effective deblurring model incorporating an <i>L</i><sub>0</sub> regularized <i>EPMG</i> prior and an <i>L</i><sub>0</sub> regularized gradient prior. Unlike previous priors, our <i>EPMG</i> prior considers both intensity and gradient information, enabling a more comprehensive distinction between clear and blurred images. Additionally, the non-overlapping patch design we adopt ensures sparsity and simplicity, significantly enhancing computational efficiency. Extensive experimental results demonstrate that our method yields superior visual quality and quantitative performance compared to several state-of-the-art methods, particularly in computational efficiency and restoration quality.</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

Enhanced Patch-Wise Maximal Gradient for Blind Image Deblurring

  • 期刊论文

  • 发表日期：

    2025年4月11日

查看全文
  • Zirui zhang,
  • Zheng Guo,
  • Zhenhua Xu,
  • Chunyong Wang,
  • Jiancheng Lai,
  • Yunjing Ji,
  • Zhenhua Li

摘要

The maximum intensity and gradient values of non-overlapping patches significantly decrease during the blurring process. To address this issue, we propose an enhanced patch-wise maximum gradient (EPMG) prior for blind image deblurring. We evaluate the statistical distribution of EPMG using a real dataset and mathematically demonstrate its effectiveness. Based on the EPMG prior, we develop an effective deblurring model incorporating an L0 regularized EPMG prior and an L0 regularized gradient prior. Unlike previous priors, our EPMG prior considers both intensity and gradient information, enabling a more comprehensive distinction between clear and blurred images. Additionally, the non-overlapping patch design we adopt ensures sparsity and simplicity, significantly enhancing computational efficiency. Extensive experimental results demonstrate that our method yields superior visual quality and quantitative performance compared to several state-of-the-art methods, particularly in computational efficiency and restoration quality.