Optimal image blending for underwater mosaics

Typical problems for creation of consistent underwater mosaic are misalignment and inhomogeneous illumination of the image frames, which causes visible seams and consequently complicates post-processing of the mosaics such as object recognition and shape extraction. Two recently developed image blending methods were explored in the literature: "gradient domain stitching" and "graph-cut" method, and they allow for improvement of illumination inconsistency and "ghosting" effects, respectively. However, due to the specifics of underwater imagery, these two methods cannot be used within a straightforward manner. In this paper, a new improved blending algorithm is proposed based on these two methods. By comparing with the previous methods from a perceptual point of view and as a potential input for pattern recognition algorithms, our results show an improvement in decreasing the mosaic degradation due to feature doubling and rapid illumination change.