DRFMamba: Multi-focus Microscopic Image Registration Fusion Based on Vision Mamba

Microscopic images provide detailed views of microstructures. They enable the observation and analysis of small biological samples, including their morphology and dynamics. However, these images often suffer from multi-focus and misalignment issues. To obtain clear and accurate images, registration and fusion are essential preprocessing steps. Currently, no unified network exists for multi-focus microscopic image registration and fusion, posing a significant challenge in extracting useful information from misaligned images. To address this challenge, we propose a high-performance model, Deep Registration Fusion Mamba (DRFMamba). The framework is divided into two stages: affine registration and fusion. In the first stage, to handle global rigid misalignment, we introduce the Deep Learning Vision Mamba Registration (DLVR) network for coarse affine registration of multi-focus microscopic images. In the DLVR network, a Linear Deformable Convolutional Vision Mamba (LDCVmamba) block is designed to capture both local and global displacement and deformation features from misaligned image pairs. This module enhances geometric disparity perception in in-focus regions while suppressing defocus interference, thereby improving affine registration accuracy. In the Deep Learning Vision Mamba Fusion (DLVF) network, a Laplacian Pyramid enhances multi-scale high-frequency representations, while the Vision Mamba (Vim) captures global structural features. Moreover, a Bidirectional Adaptive Cross-Alignment Fusion (ACAF) module achieves pixel-level feature alignment between image pairs, enabling bidirectional exchange of complementary information and adaptive fusion, further enhancing overall fusion performance. Extensive experimental results demonstrate that DRFMamba exhibits strong registration and fusion capabilities on multi-focus image datasets.