.png){kind=icon}
.png){kind=icon}
Return
Hybrid transformer–CNN framework for light field angular super-resolution
.png){kind=icon}
PRE
AI
DOI：10.1016/j.imavis.2026.106176.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
•
Hybrid Feature Extraction Block (Hybrid _FEB): We propose a novel Hybrid Feature Etraction Block that performs sequential multi- representation feature learning for LFASR. The input features are first transformed into horizontal and vertical EPI representations, where a shared Transformer captures long range spatial-angular dependencies along epipolar lines. The enhanced features are then reformulated into the MacroPI representation, enabling dedicated spatial and angular CNN branches to extract localized correlations. By bridging EPI-based global dependency modeling with MacroPI-based local feature extraction, the proposed Hybrid_FEB effectively captures both long-range geometric relationships and localized spatial-angular correlations for accurate LF reconstruction.
•
Hierarchical Feature Refinement: A series of Hybrid _FEB modules are hierarchically cascaded to progressively refine spatial–angular representations. The outputs from these stages are then fused and further enhanced through a spatial residual block, enabling effective multilevel feature integration. This hierarchical refinement design allows the network to capture long-range view interactions while reinforcing fine-grained geometric and structural details, thereby achieving high-fidelity light field reconstruction even in scenes with wide disparity variations.
•
Robust Performance Across Datasets: Extensive experiments on both synthetic and real-world LF datasets demonstrate that the proposed method consistently achieves strong quantitative performance. The framework effectively preserves epipolar consistency, structural integrity, and angular coherence, validating its reliability and generalizability in diverse reconstruction scenarios.
Journal
IF:
4.2
Papers:
4.0K
Citations:
6.7K