Due to an enormous production of multimedia materials, a mosaic has become a valuable representation of the video sequences in the application field such as video surveillance, interactive video edition and manipulation, low-bitrate video transmission, and so on. These applications require efficient representations of huge video data and suitable methods of storing and accessing the video data. In this thesis, we propose the efficient alignment algorithm to compensate for small amounts of motion parallax introduced by unmodeled distortions. Also, we propose a new type of mosaic to minimize the loss of image data occurred by coordinate transformations.
After the global alignment has been run, there may still be localized mis-registrations present in the mosaic, due to deviations from the idealized parallax-free motion model. To estimate the residual motion vectors occurred by mis-registrations consumes heavy computational powers. In the proposed algorithm, residual motion vectors are estimated with the adequate size of estimation and measurement windows and with adjustment of initial vectors according to the established priority. By construction of mosaic with the proposed algorithm, we demonstrate the removal of seams by mis-registrations.
While many researchers have focused on the issue of how to develop an exact representation with algorithms related to image alignment and scene composition, little attention has been attracted on the issue of how to minimize the loss of image data occurred by coordinate transformations in the construction of a single-view mosaic. This kind of loss occurs in some regions of a single-view mosaic, where sufficient samples are not assigned to represent the image data contained in high resolution frames before transformations.
A resolution map is introduced in order to measure the spatially required number of samples. By segmenting it, we decide which regions will be included and what will be the representative of the resolution in the considered layer. Using the representatives, each layer has the specific coordinate system and intensities of all the samples are blended with the recalculated transforming parameters. We also propose an efficient coding scheme for the proposed mosaic. Simulation results show the increase of the image qualities of the frames reconstructed from the proposed mosaic with given storage volumes.