Two-Phase Multimodal Image Fusion Using Convolutional Neural Networks

The fusion of multiple imaging modalities presents an important contribution to machine vision but remains an ongoing challenge due to the limitations in traditional calibration methods that perform a single, global alignment. For depth and thermal imaging devices, sensor and lens intrinsics (FOV, resolution, etc.) may vary considerably, making per-pixel fusion accuracy difficult. In this paper, we present AccuFusion, a two-phase non-linear registration method to fuse multimodal images at a per-pixel level to obtain an efficient and accurate image registration. The two phases: the Coarse Fusion Network (CFN) and Refining Fusion Network (RFN), are designed to learn a robust image-space fusion that provides a non-linear mapping for accurate alignment. By employing the refinement process, we obtain per-pixel displacements to minimize local alignment errors and observe an increase of 18% in average accuracy over global registration.