Building Panoramic Image Mosaics of Coral Reefs
MetadataShow full item record
Panoramic image mosaic is an important tool for accessing the preserving condition of coral reefs sites. In recent years, a number of methods and applications are made to create panoramic image mosaics by stitching a sequence of images together to get a wider field of view. However, it is not appropriate to apply existing methods and applications to our dataset directly. The main reason is that our underwater videos are acquired not by underwater machines, but coral-conservation researchers. In this case, the stitching process becomes more challenging due to a number of problems: illumination changes, large parallax, water drift, as well as underwater moving objects. To create high-quality mosaics by stitching thousands of images extracted from multiple videos of a survey site, algorithms need to recover the positional information of each image and handle above problems. In this dissertation, we combine several novel techniques to stitch image sequences from multiple videos as-rigid-as-possible and to create image mosaics as-natural-as-possible. To stitch two images with large parallax, we propose to use two novel parallax-minimization stitching algorithms. The first algorithm is based on superpixel segmentation. It separates the source image into parallax region and non-parallax region, and applies different transformations to each region to align with the target image. The second algorithm uses a modified optimal seam to decide which pixel to choose from each image, followed by a modified Poisson blending method to improve quality. The modified cost function guides the seam goes through well-aligned non-parallax background region and preserve the shape of the foreground objects. By comparing to existing methods and software, we show that our parallax-minimization stitching algorithms can significantly reduce ghosting and distortions. To create mosaic of a single video, we need to recover the camera motion trajectory first to obtain positional information of each image. We propose to use a feature-tracking algorithm to find dominant motion between time-consecutive images. Here, we denote images as nodes and dominant motion as link. Images are connected sequentially based on the dominant motions to form a topological graph. We then use a graphical strategy to detect and remove low-quality region images caused by water drift or moving objects. The remaining connection images from the low-quality regions are stitched using the proposed parallax-tolerant algorithms and the rest images are stitched using Affine transformation. To create mosaic of multiple videos, we first use a modified all-to-all matching algorithm to find connection regions between adjacent videos. We stitch all the images together based on the connection regions. We then propose to use an extended optimal seam to improve the mosaic quality. The optimal seam is composed of a number of segments computed using different cost functions and different algorithms. The stitched mosaic, which has a visible seam along overlapping region, is finally blended using the modified Poisson blending method to smooth the color differences. We compare our contributions to the state-of-the-art. Our experiments demonstrate that our method efficiently creates high-quality image mosaics with minimal distortions and minimal ghosting.