Occlusion artifacts significantly hinder light field (LF) image reconstruction, especially in complex scenes. We propose a spectral normalized U-Net for LF occlusion removal, which begins by stacking LF views and extracting view-dependent features using a local feature encoder. To capture spatial complexity, ResASPP enable multi-scale context aggregation, while channel attention enhances occlusion-related features. Spectral normalization is applied to all convolutional layers to improve training stability and generalization. The encoder-decoder structure with skip connections preserves fine details. Experimental results show our method restores occluded regions more accurately than baselines.

Occlusion removal in light-field images remains a significant challenge, particularly when dealing with large occlusions. An architecture based on end-to-end learning is proposed to address this challenge that interactively combines CSPDarknet53 and the bidirectional feature pyramid network for efficient light-field occlusion removal. CSPDarknet53 acts as the backbone, providing robust and rich feature extraction across multiple scales, while the bidirectional feature pyramid network enhances comprehensive feature integration through an advanced multi-scale fusion mechanism. To preserve efficiency without sacrificing the quality of the extracted feature, our model uses separable convolutional blocks. A simple refinement module based on half-instance initialization blocks is integrated to explore the local details and global structures. The network’s multi-perspective approach guarantees almost total occlusion removal, enabling it to handle occlusions of varying sizes or complexity. Numerous experiments were run on sparse and dense datasets with varying degrees of occlusion severity in order to assess the performance. Significant advancements over the current cutting-edge techniques are shown in the findings for the sparse dataset, while competitive results are obtained for the dense dataset.