文献类型：期刊文献

英文题名：Multi-focus image fusion based on multi-scale feature extraction and edge preservation techniques

作者：Zhao, Baojun[1];Luo, Fei[1,2];Pino, Luis Rojas[3];Ding, Weichao[1];Zhang, Xueqin[1,2]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China;[2]Shanghai Dev Ctr Comp Software Technol, Shanghai Key Lab Comp Software Evaluating & Testin, Shanghai 201203, Peoples R China;[3]Univ San Sebastian, Sch Engn, Santiago 8320000, Chile

年份：2026

卷号：341

外文期刊名：KNOWLEDGE-BASED SYSTEMS

收录：;EI(收录号：20261320380734);WOS:【SCI-EXPANDED(收录号:WOS:001737710200002)】；

基金：This research was supported by National Key Research and Development Program of China (2024YFC3307700) , Nature Science Foundation of Shanghai (22ZR1416500) , Shanghai Basic Research Special Zone Program (22TQ1400100-16) .

语种：英文

外文关键词：Multi-focus image fusion; Unsupervised learning; Multi-scale feature extraction; Edge-preserving filtering; Manifold theory

摘要：As a popular image enhancement technique, decision map-based multi-focus image fusion (MFIF) integrates partially focused images into fully focused composites without the problem of image distortion. However, conventional decision map-based methods face dual limitations: (1) Encoders in existing unsupervised methods cannot fully capture contextual information, which weakens the robustness of activity-level measurement. (2) Post-processing techniques lose edge consistency between decision maps and source images, caused by defocus spread effects at focus boundaries. To address those issues, we propose EP-MFIF, an unsupervised framework that synergizes multi-scale contextual modeling and geometry-aware edge preservation. First, a multi-branch convolution-based encoder with triple-branch receptive fields (5 & times;5, 9 & times;9, 13 & times;13) is proposed to capture multi-scale contextual features, enabling robust activity-level measurement while preserving spatial resolution. Second, as our core contribution, the post-processing is designed as a geometry-aware, manifold-based refinement module that redefines the distance metric (curvature-consistent three-point geodesic) and performs an isometric 1D domain transformation, enabling recursive joint filtering that tightly adheres to focus boundaries and outperforms guided/bilateral variants in MFIF. Third, boundary-consistent focus and defocus masks are generated via a concise two-stage fusion coupling spatial-frequency activity estimation with edge-preserving refinement. Evaluated on the Lytro dataset, EP-MFIF achieves state-of-the-art performance, improving the Qmi metric by 2.8% over existing unsupervised methods. The methodology effectively bridges multi-scale feature integration with edge-preserving theory, offering practical value for applications requiring precise focus boundary determination in microscopy, computational photography, and optical inspection systems.