EDFusion: Edge-guided attention and dynamic receptive field with dense residual for multi-focus image fusion

Abstract

Multi-focus image fusion (MFIF) synthesizes a fully focused image by integrating multiple partially focused images captured at distinct focal planes of the same scene. However, existing methods often fall short in preserving edge and texture details. To address this issue, this paper proposes a network for multi-focus image fusion that incorporates edge-guided attention and dynamic receptive field dense residuals. The network employs a specially designed dynamic receptive field dense residual block (DRF-DRB) to achieve adaptive multi-scale feature extraction, providing rich contextual information for subsequent fine fusion. Building on this, an edge-guided fusion module (EGFM) explicitly leverages the differences in source images as edge priors to generate dedicated weight maps for each feature channel, enabling precise boundary preservation. To efficiently model global dependencies, we introduce a multi-scale token mixing transformer (MSTM-Transformer), designed to reduce computational complexity while enhancing cross-scale semantic interactions. Finally, a refined multi-scale context upsampling module (MSCU) reconstructs high-frequency details. Experiments were conducted on five public datasets, comparing against twelve state-of-the-art methods and evaluated using nine metrics. Both quantitative and qualitative results demonstrate that the proposed method significantly outperforms existing approaches in fusion performance. Notably, on the Lytro dataset, the proposed method ranked first across eight core metrics, achieving high scores of 1.1946 in the information preservation metric ($Q_{NMI}$) and 0.7629 in the edge information fidelity metric ($Q_{AB/F}$).