Robust patchmatch HDR image reconstruction for deghosting

Abstract

• We propose an optimization function to search the optimal patch. • We design a registration scheme to align motion of source images. • We adopt a effective deghosting solution consisting of registration and matching. Aiming at the deghosting problem of High Dynamic Range (HDR) imaging technology, different from the existing algorithms of simply eliminating moving objects or only using a separate pre-registration based on homography transformation or optical flow, we propose an integrated HDR imaging scheme including Registration and Matching based on Patch (HDR-RMP) for deghosting in complex dynamic scenes. In our scheme, the combination of Affine Transformation (AT) and Normalization Mutual Information (NMI) has an implicit registration effect on the input images. Furthermore, NMI and Structural Similarity Index Measurement (SSIM) are fit to establish the final optimal function including the strategies of registration and matching, and the registration accuracy is promoted by this way, thereby the ability to remove ghosts is improved and the time complexity is reduced simultaneously. Finally, the accuracy of image matching is enhanced. The results show that the deghosting effect of HDR-RMP algorithm is remarkable, especially in dynamic scenes with large and mass moving objects, which not only improves the fusion quality, but also reduces the algorithmic complexity.