The end-to-end approach that directly learns the mapping from multi-focus images to fused images has been widely used recently, which achieves excellent performance in dealing with complex scenes. However, the fusion quality of this approach falls short of decision map-based methods, as this approach can preserve the original pixels of the focused regions in the fused image, while end-to-end methods use network inference results with pixel-wise regression errors, resulting in low fidelity of the fused images. To mitigate this limitation, we propose SwinMFF, which effectively captures long-range dependencies across the source images via the swin transformer to reduce pixel-wise regression errors, achieving high-fidelity end-to-end fusion while simultaneously alleviating edge artifacts in the fused image. Extensive experiments demonstrate that SwinMFF outperforms the other 28 state-of-the-art methods in both subjective visual quality and quantitative metrics. The codes are available at https://github.com/Xinzhe99/SwinMFF.
In recent years, deep learning-based multi-focus image fusion (MFF) methods have demonstrated remarkable performance. However, their reliance on complex network architectures often demands substantial computational resources, limiting practical applications. To address this, we propose LightMFF, an ultra-lightweight fusion network that achieves superior performance with minimal computational overhead. Our core insight is to reformulate the multi-focus fusion problem from a classification perspective to a refinement perspective, where coarse initial decision maps and explicit edge information are leveraged to guide the final decision map generation. This novel formulation enables a significantly simplified architecture, requiring only 0.02 M parameters while maintaining state-of-the-art fusion quality. Extensive experiments demonstrate that LightMFF achieves real-time performance at 0.02 s per image pair with merely 0.06 G FLOPs, representing a 98.05% reduction in computational cost compared to prior approaches. Crucially, LightMFF consistently surpasses existing methods across standard fusion quality metrics.
12 March 2025 in Computers and Electrical Engineering
Limited depth of field in modern optical imaging systems often results in partially focused images. Multi-focus image fusion (MFF) addresses this by synthesizing an all-in-focus image from multiple source images captured at different focal planes. While deep learning-based MFF methods have shown promising results, existing approaches face significant challenges. Convolutional Neural Networks (CNNs) often struggle to capture long-range dependencies effectively, while Transformer and Mamba-based architectures, despite their strengths, suffer from high computational costs and rigid input size constraints, frequently necessitating patch-wise fusion during inference—a compromise that undermines the realization of a true global receptive field. To overcome these limitations, we propose MLP-MFF, a novel lightweight, end-to-end MFF network built upon the Pyramid Fusion Multi-Layer Perceptron (PFMLP) architecture. MLP-MFF is specifically designed to handle flexible input scales, efficiently learn multi-scale feature representations, and capture critical long-range dependencies. Furthermore, we introduce a Dual-Path Adaptive Multi-scale Feature-Fusion Module based on Hybrid Attention (DAMFFM-HA), which adaptively integrates hybrid attention mechanisms and allocates weights to optimally fuse multi-scale features, thereby significantly enhancing fusion performance. Extensive experiments on public multi-focus image datasets demonstrate that our proposed MLP-MFF achieves competitive, and often superior, fusion quality compared to current state-of-the-art MFF methods, all while maintaining a lightweight and efficient architecture.