Parallax Image Stitching Based on Optical Flow and Secondary Projection Correction

Parallax image stitching often encounters issues such as misalignment of internal objects within the images and the presence of artifacts in the stitched images. To address these challenges, this paper proposes a parallax image stitching method based on optical flow and secondary projection correction. In terms of image registration, a multi-layer feature constrained homography estimation network is constructed. This network employs an iterative depth-wise approximation technique to accurately predict the offset between the images to be registered. Additionally, a dataset of disparity images is constructed to facilitate the training of the network. During the image fusion stage, the concept of optical flow interpolation is employed to blend the overlapping regions of the images. Subsequently, a rematching process is applied to both the overlapping and non-overlapping regions to calculate the homography matrix, which is then utilized for secondary projection transformation of the non-overlapping regions. Experimental results demonstrate that the proposed method achieves visually pleasing parallax image stitching outcomes.