Revolutionizing Occlusion Removal in Light-Field Images: Introducing TriORU^2-Net++
Introducing \\(\\hbox {TriORU}^2\\)-Net++, a cutting-edge architecture tailored to tackle occlusion removal challenges in light-field (LF) images. This innovative model employs adaptive attention-guided feature integration and progressive hierarchical reconstruction to effectively address occlusion restoration issues. By incorporating a ResASPP-AttFPN feature extractor, the model can fuse multiscale features and emphasize spatial cues crucial for occlusion localization. The core of the framework is a tri-stage \\(\\hbox {U}^2\\)-Net++ reconstruction module that progressively restores occluded regions through three interconnected encoder-decoder stages of varying depth. To enhance detail preservation and structural consistency, a residual feature refiner (RFR) is introduced to sharpen object boundaries. Extensive evaluations demonstrate that the proposed method outperforms existing LF occlusion removal techniques in both quantitative metrics and visual quality, making it a valuable tool for preprocessing large-scale visual datasets in the Big Data era. This work was supported by various grants and programs, highlighting the collaborative effort behind this research.