Locally Aligned Image Stitching Based on Multi-Feature and Super-Pixel Segmentation With Plane Protection
Jialiang Li, Dong Mei Wu, Pinqun Jiang, Zili Li, Shuxiang Song · IEEE Access · 2021
The number and accuracy of image feature matching directly affect the accuracy of image warping, which is gaining widespread attention in image stitching. From a single plane to a regular multi-plane warping model, the alignment accuracy is gradually improved. To further improve the accuracy of image alignment, this paper proposes a multi-plane alignment method based on superpixel segmentation, which preserves the integrity of local planes as much as possible to reduce ghosting. Recent warps prove that line features provide strong correspondences, especially in low-textured cases. Moreover, image segmentation methods such as superpixel segmentation have the function of protecting the integrity of the object. On the one hand, in the feature matching stage, our approach introduces superpixel segmentation to refine the matching. The homography combines line features to enrich as many matching points as possible. On the other hand, to solve the problem of misalignment caused by local warping, the proposed method makes full use of the characteristics of superpixels to perform irregular plane segmentation to avoid the traditional rectangular segmentation method from dividing different planes into the same grid. Experimental results demonstrate that the proposed method outperforms some state-of-the-art warps from both qualitative and quantitative aspects, including the as-projective-as-possible warp (APAP), the as-natural-as-possible warp (ANAP), the global similarity prior (GSP), and so on.