Focus-Aware Fusion for Enhanced All-in-Focus Light Field Image Generation

All-in-focus images, containing comprehensive clear scene information, serve as the cornerstone for numerous high-precision light field (LF) measurement applications. However, generating all-in-focus images from original LF data necessitates exploiting the scene depth via refocusing techniques. Yet, in real-world scenarios, the scene depth remains largely unknown. Addressing this challenge, this article proposes a focus-aware fusion method based on focal stack images to generate all-in-focus LF images devoid of prior depth information. This method fully harnesses the diverse focus information present in multifocus LF images. Initially, we employ refocusing techniques to produce a sequence of continuously refocused images at arbitrary distances. In addition, we introduce a grouping strategy to categorize focal stack images with significant disparities into a foreground group and a background group, ensuring seamless depth transitions within each group. Subsequently, multidirectional log-Gabor filtering is utilized to fuse images within each group. Finally, the resulting two images are merged to yield the precise all-in-focus image. Throughout the fusion process, we introduce a multistage precise focus decision map generation strategy to achieve accurate detection of focus regions, while neighbor distance filtering is applied to retain detail information and enhance edge accuracy. Extensive experiments on the simulated and real-world LF datasets demonstrate superior performance over ten mainstream fusion methods in terms of robustness, edge preservation, detail clarity, and visual perception. Our source code is made available at https://github.com/banxq/FocusLF