FIIF-Net: an end-to-end multi-focus image fusion framework for cosmic dust microscopic imaging

Abstract

Characterizing the microscopic morphology of cosmic dust particles is essential for understanding planetary formation and evolution; however, the shallow depth of field inherent in miniature imaging systems precludes single-shot all-in-focus acquisition. Multi-focus image fusion (MFIF) computationally extends the depth of field, yet existing methods degrade markedly under spatial misalignment and produce hard-boundary stitching artifacts at focus transition regions. We present FIIF-Net, an end-to-end framework designed for misaligned multi-focus microscopic dust images. In the alignment stage, a residual feature downsampling pyramid equipped with a learnable repair factor recovers fine morphological textures lost during conventional downsampling of low-contrast microscopic features, and a dynamic lookup operator adaptively modulates the correlation search radius to accommodate vibration-induced, spatially non-uniform displacements. In the fusion stage, a dual-path interactive fusion strategy incorporating STA bidirectionally couples the image-content path with the focus-guidance path, enabling content features and continuous focus cues to mutually modulate each other and thereby replacing binary decision-map stitching with physically consistent soft transitions. Zero-shot super-resolution is further integrated to compensate for the edge attenuation inherent in defocus-aware blending. A dedicated multi-focus misaligned cosmic dust particle dataset, combining real focal stacks with physics-based rendered data, is constructed for training and evaluation. Experiments on both synthetic and real data demonstrate that FIIF-Net surpasses state-of-the-art MFIF methods quantitatively and visually, particularly in preserving edge details within focal transition regions.