DAORNet: Dense-attentive network for occlusion removal in light field images

Abstract

A light field (LF) image captures abundant spatial–angular information across multiple viewpoints, offering rich cues for occlusion removal; however, it also presents significant challenges due to the discontinuities introduced by complex occluded regions. This paper introduces a dense-attentive occlusion removal network (DAORNet), an end-to-end architecture that combines multi-scale feature extraction, attention-driven occlusion reconstruction, and a novel dense refinement design to restore occluded content with high structural and perceptual accuracy. Specifically, DAORNet consists of three main components. First, a spatial pyramid pooling fast (SPPF)-enhanced ResNet18 (SPPF-ResNet18) serves as the feature extractor, capturing hierarchical spatial–angular correlations across multiple receptive fields. Second, a coordinate attention-augmented BiFPN (CoordAtt-BiFPN) acts as the occlusion reconstruction (OR) module, enabling adaptive multi-scale fusion while maintaining geometric consistency and contextual dependencies across views. Finally, a refinement module incorporating two cascaded, newly designed DenseRefiners (DRs) and an OutBlock, which focus on texture completion, enhances structural consistency and high-frequency detail recovery. Each DR employs dual-branch processing to aggregate contextual and multi-dilation features, expanding receptive coverage and strengthening contextual perception. Attention mechanisms further support effective detail preservation, improving texture continuity and edge sharpness. Experimental evaluations on standard LF datasets demonstrate that DAORNet achieves significant improvements in PSNR and SSIM over state-of-the-art (SOTA) methods, ultimately yielding visually consistent, occlusion-free LF reconstructions.