Scholarlink

<p>We introduce <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(\hbox {TriORU}^2\)</EquationSource></InlineEquation>-Net++, a novel three-stage architecture designed to address the persistent challenge of occlusion removal in light-field (LF) images by leveraging adaptive attention-guided feature integration and progressive hierarchical reconstruction. Unlike existing methods that struggle to fully exploit spatial hierarchies and adaptively restore occluded regions across scales, our model incorporates a ResASPP-AttFPN feature extractor, which integrates Residual Atrous Spatial Pyramid Pooling (ResASPP) with a spatial attention-enhanced Feature Pyramid Network (AttFPN) to selectively fuse multiscale features while emphasizing salient spatial cues essential for occlusion localization. The core of our framework is a tri-stage <InlineEquation ID="IEq4"><EquationSource Format="TEX">\(\hbox {U}^2\)</EquationSource></InlineEquation>-Net++ reconstruction module, which performs progressive restoration through three hierarchically connected encoder-decoder stages of decreasing depth (4-level, 3-level, and 2-level), each built on VGG-based blocks and dense skip connections to recover increasingly refined background content. To further enhance detail preservation and structural consistency, we introduce a residual feature refiner (RFR) that consolidates residual cues and sharpens the boundaries of objects. Extensive experimental evaluations demonstrate that the proposed method surpasses recent state-of-the-art (SOTA) LF occlusion removal approaches—representing the most advanced and best-performing techniques reported in the literature—in both quantitative metrics and visual reconstruction quality. Specifically, our model achieves average improvements of 0.86&#xa0;dB in PSNR and 0.016 in SSIM across real-world (CD scene) and synthetic LF datasets, including sparse (4-Syn, 9-Syn) and dense (Single-Occ, Double-Occ) settings. This capability is particularly relevant to the Big Data paradigm, where large-scale visual datasets demand robust preprocessing to remove occlusions and ensure reliable downstream analytics. By improving LF data fidelity while remaining efficient, our model supports scalable pipelines for high-volume visual data processing.&#xa0;&#xa0;&#xa0;&#xa0;</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

TriORU2-Net++: attention-guided three-stage U2-Net++ for light field occlusion removal

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月1日

查看全文
  • Mostafa Farouk Senussi,
  • Mahmoud Abdalla,
  • Mahmoud SalahEldin Kasem,
  • Mohamed Mahmoud,
  • Hyun-Soo Kang

摘要

We introduce \(\hbox {TriORU}^2\)-Net++, a novel three-stage architecture designed to address the persistent challenge of occlusion removal in light-field (LF) images by leveraging adaptive attention-guided feature integration and progressive hierarchical reconstruction. Unlike existing methods that struggle to fully exploit spatial hierarchies and adaptively restore occluded regions across scales, our model incorporates a ResASPP-AttFPN feature extractor, which integrates Residual Atrous Spatial Pyramid Pooling (ResASPP) with a spatial attention-enhanced Feature Pyramid Network (AttFPN) to selectively fuse multiscale features while emphasizing salient spatial cues essential for occlusion localization. The core of our framework is a tri-stage \(\hbox {U}^2\)-Net++ reconstruction module, which performs progressive restoration through three hierarchically connected encoder-decoder stages of decreasing depth (4-level, 3-level, and 2-level), each built on VGG-based blocks and dense skip connections to recover increasingly refined background content. To further enhance detail preservation and structural consistency, we introduce a residual feature refiner (RFR) that consolidates residual cues and sharpens the boundaries of objects. Extensive experimental evaluations demonstrate that the proposed method surpasses recent state-of-the-art (SOTA) LF occlusion removal approaches—representing the most advanced and best-performing techniques reported in the literature—in both quantitative metrics and visual reconstruction quality. Specifically, our model achieves average improvements of 0.86 dB in PSNR and 0.016 in SSIM across real-world (CD scene) and synthetic LF datasets, including sparse (4-Syn, 9-Syn) and dense (Single-Occ, Double-Occ) settings. This capability is particularly relevant to the Big Data paradigm, where large-scale visual datasets demand robust preprocessing to remove occlusions and ensure reliable downstream analytics. By improving LF data fidelity while remaining efficient, our model supports scalable pipelines for high-volume visual data processing.