Scholarlink

<p>This study introduces a novel image restoration framework that integrates parameter-assisted inpainting and edge-based inpainting algorithms with the advanced U-Net Mamba network architecture. The proposed hybrid approach leverages structural priors from traditional methods alongside the contextual learning capabilities of modern deep learning networks, emphasizing its novelty in enhancing both computational efficiency and image restoration accuracy. Key findings highlight the model’s ability to achieve high-quality reconstructions while maintaining resource efficiency. The VMI model was trained for 400 epochs using the CLIC dataset (1287 images: 1048 training, 61 validations, 178 testing) and the KODAK dataset (1803 images: 1442 training, 144 validations, 217 testing). Images were resized to 256 × 256 pixels for computational efficiency. A cyclical learning rate strategy facilitated model optimization, as shown in the loss, PSNR, and SSIM trends. Experimental results demonstrate the model’s robust inpainting capabilities, achieving effective convergence with a final loss of approximately 0.3, an SSIM of 0.8, and a PSNR of 28&#xa0;dB. The KODAK dataset outperformed CLIC in reconstruction quality, achieving higher PSNR and SSIM values, though with slightly higher loss. These outcomes indicate superior reconstruction quality, effectively restoring missing regions while preserving structural and perceptual fidelity. Visual analysis further confirms the model’s ability to maintain structural integrity and perceptual quality in inpainted images. This research establishes a compact yet powerful framework for image restoration, offering a highly precise and resource-efficient solution. The findings set new benchmarks for applications requiring high accuracy and efficiency, emphasizing the novelty of integrating traditional structural priors with deep learning methodologies to enhance image restoration performance.</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

Enhancing image restoration: parameter-assisted and edge-based inpainting with U-Net mamba

  • 期刊论文

  • 发表日期：

    2025年4月5日

查看全文
  • Afshin Sandooghdar,
  • Farzin Yaghmaee

摘要

This study introduces a novel image restoration framework that integrates parameter-assisted inpainting and edge-based inpainting algorithms with the advanced U-Net Mamba network architecture. The proposed hybrid approach leverages structural priors from traditional methods alongside the contextual learning capabilities of modern deep learning networks, emphasizing its novelty in enhancing both computational efficiency and image restoration accuracy. Key findings highlight the model’s ability to achieve high-quality reconstructions while maintaining resource efficiency. The VMI model was trained for 400 epochs using the CLIC dataset (1287 images: 1048 training, 61 validations, 178 testing) and the KODAK dataset (1803 images: 1442 training, 144 validations, 217 testing). Images were resized to 256 × 256 pixels for computational efficiency. A cyclical learning rate strategy facilitated model optimization, as shown in the loss, PSNR, and SSIM trends. Experimental results demonstrate the model’s robust inpainting capabilities, achieving effective convergence with a final loss of approximately 0.3, an SSIM of 0.8, and a PSNR of 28 dB. The KODAK dataset outperformed CLIC in reconstruction quality, achieving higher PSNR and SSIM values, though with slightly higher loss. These outcomes indicate superior reconstruction quality, effectively restoring missing regions while preserving structural and perceptual fidelity. Visual analysis further confirms the model’s ability to maintain structural integrity and perceptual quality in inpainted images. This research establishes a compact yet powerful framework for image restoration, offering a highly precise and resource-efficient solution. The findings set new benchmarks for applications requiring high accuracy and efficiency, emphasizing the novelty of integrating traditional structural priors with deep learning methodologies to enhance image restoration performance.