Scholarlink

We propose a lightweight image inpainting framework that couples a Radial Basis Function Neural Network (RBFNN) with a D2Q9 Lattice Boltzmann Method (LBM). The model learns a mass-conserving collision update \(\varOmega \) inside holes and a mask probability per pixel, while a strict variant guarantees no change outside the mask. On grayscale Caltech101 with diverse masks, our method yields coherent, edge-preserving reconstructions and outperforms a non-LBM RBF patch regressor on medium and large holes. We report PSNR, SSIM and hole-only PSNR, with ablations on RBF centers, relaxation, and post-processing.

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

Radial Basis Function Neural Networks for Collision Learning and Mask Detection in Image Inpainting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yassine Douich,
  • Hassan Silkan,
  • Youssef Hanyf

摘要

We propose a lightweight image inpainting framework that couples a Radial Basis Function Neural Network (RBFNN) with a D2Q9 Lattice Boltzmann Method (LBM). The model learns a mass-conserving collision update \(\varOmega \) inside holes and a mask probability per pixel, while a strict variant guarantees no change outside the mask. On grayscale Caltech101 with diverse masks, our method yields coherent, edge-preserving reconstructions and outperforms a non-LBM RBF patch regressor on medium and large holes. We report PSNR, SSIM and hole-only PSNR, with ablations on RBF centers, relaxation, and post-processing.