Atom Search–Sine Cosine Algorithm Guided Lightweight CNN for Multi-Focus Image Fusion with Enhanced Edge Preservation and Information Transfer

Multi-focus image fusion addresses the inherent depth-of-field limitation of optical systems by combining partially focused source images into a single all-in-focus composite image. This paper presents a two-stage hybrid fusion framework, the first stage employs a weighted averaging strategy, where the fusion weights are optimized using the proposed Atom Search–Sine Cosine Algorithm (ASSCA). The second stage refines the initial fusion result using a lightweight deep convolutional neural network (DCNN), with its optimization parameters adjusted using the same ASSCA strategy. By integrating the exploratory characteristics of Atom Search with the directional exploitation of Sine-Cosine updates, ASSCA balances global search and local refinement to improve fusion quality and reduce artifacts. Evaluated on the complete Lytro dataset (20 image pairs), the proposed ASSCA–DCNN method achieves a mean Mutual Information of 1.46 ± 0.03, an edge-based similarity (Q_AB/F) of 0.75 ± 0.01 (ranging from 0.74 to 0.76), entropy exceeding 7.1, and noticeably improved spatial frequency with strong preservation of sharp boundaries. These results are comparable with recent lightweight hybrid methods and approach the performance of state-of-the-art transformer-based techniques, such as LightMFF (Q_AB/F ≈ 0.7588), while requiring fewer parameters and avoiding GPU acceleration during inference. Visual evaluations indicate reduced halo artifacts and improved contrast preservation, particularly in textured regions. The approach provides an interpretable and computationally efficient solution suitable for embedded vision applications. Statistical evaluation across all 20 Lytro image pairs reports a mean PSNR of 28.4 ± 2.3 dB and RMSE of 24.8 ± 3.1. The optimization process converges within 150 iterations across the evaluated cases. Overall, the proposed method provides a practical, low-resource alternative with performance comparable to lightweight transformer models. The proposed framework aligns with UN Sustainable Development Goal 9 (Industry, Innovation and Infrastructure) by contributing to efficient, lightweight image processing techniques for resource-limited and sustainable imaging systems.