MDFAT: Interactive mask decoupling and frequency-adaptive transformer for multi-focus image fusion

Multi-focus image fusion aims to combine multiple images of the same scene, each with different focused regions, into a single globally clear image. Existing methods still face challenges in effectively modeling global context and accurately separating complementary focused features. To address this, this paper proposes a network model that integrates a fusion interactive mask decoupling mechanism with a frequency-adaptive transformer dual mechanism. By migrating attention computation to the frequency domain, we design a frequency-adaptive transformer mechanism with adaptive blocking that dynamically partitions feature maps into optimal blocks for frequency-domain analysis. Utilizing the Fourier transform, this approach converts the traditionally high-complexity matrix operations in the spatial domain into efficient element-wise multiplications, significantly reducing the computational cost of modeling long-range dependencies. Meanwhile, by introducing a dynamic mask-guided feature decoupling strategy, the feature flow is explicitly divided into three independent paths targeting the foreground, background, and global context, respectively, enabling precise selection and targeted fusion of the clear regions from the two source images. We conducted systematic experiments on four standard datasets and performed comprehensive comparisons with twelve state-of-the-art algorithms. Experimental results demonstrate that the proposed method outperforms existing approaches in all objective evaluation metrics and subjective visual quality, exhibiting superior overall performance.