Robust Underwater Image Stitching Based on Graph Matching

Xu Yang, Zhi-Yong Liu, Chuan Li, Jing-Jing Wang, Hong Qiao

Abstract: Image stitching is important in intelligent perception and manipulation of underwater robots. In spite of a well developed technique, it is still challenging for underwater images because of their inevitable appearance ambiguity. For the feature based underwater image stitching, robust feature correspondence is the key because most other algorithmic parts are less directly associated with the characteristics of underwater images. Structural information between feature points may be helpful for robust feature correspondence, and based on this idea the paper proposes a robust underwater image stitching method by incorporating structural cues as additional information, whose effectiveness is validated on real underwater images. Specifically, the appearance information and structural cues are integrated by a labeled weighted graph, and the underwater image correspondence is formulated by graph matching. After geometric transformation estimation, the underwater images are finally blended into a wider viewing image.