Research On Reduction Algorithms In Digital Multimedia

Reduction algorithms in digital multimedia focus on researches on how to reduce da-ta scales and how to simplify data processing methods, while preserving intrinsic proper-ties of digital multimedia data. The key point of reduction algorithms in digital multime-dia is how to utilize the reduced data or method to better represent intrinsic information of the original data, which usually employs greedy algorithms or iterative algorithms to control the reduced size and the approximation accuracy. We have made further research-es on four reduction issues, which are related to digital image, digital video and digital geometry model, and we have obtained innovative achievements as follows:Firstly, we have presented a novel algorithm for realtime content-aware image re-sizing. The resizing is performed by warping a triangular mesh over the image, which captures the image feature and saliency information. The horizontal and vertical scales of all triangles and coordinates of all vertices could be simultaneously obtained by min-imizing a quadratic energy function, which is related to a sparse linear system. The algorithm is able to prevent the main objects form severe distortions, including feature lines and feature curves in the image, during resizing. Moreover, the quadratic optimiza-tion could be promoted to a quadratic programming to strictly prevent self-intersections in the result image. Experimental results have been tested to clarify the efficiency and the effectiveness of the algorithm. The algorithm can be performed in an interactive rate and outperforms previous approaches.Secondly, we have proposed a transformed representation technique for digital video. We have proposed the concept of video polyline, which is an effective key frame extrac-tion and navigation tool for videos. The high dimensional video polyline is constructed by detecting and clustering video features, and the video polyline is to be mapped in-to the3D space for the purpose of visualization. The multi-level representation of video polyline is constructed by an iterative polyline simplification method respect to geometric metrics, and in a definite level, each node of the polyline represents the key frame of the video. Video polyline provides key frame information and intrinsic video information. Moreover, users are able to effectively and efficiently navigate the video by using video polyline instead of timeline bar.After that, we have proposed an improved algorithm for variational mesh approxi-mation, which is an efficient algorithm for digital geometry mesh simplification based on greedy algorithm. After the facet number of the target mesh has been set, we use the nor-mal related energy function to measure the approximation accuracy. We have computed the discrete formulation of the energy function and designed a methodology for minimiz-ing the energy function with high efficiency based on a greedy merging algorithm. The method also controls the distribution of facets of target mesh based on curvature infor-mation of the original mesh, which achieves better quality of target mesh. The proposed algorithm is efficient and intuitive in geometric view, which is valuable to be integrated into digital geometry modeling systems.Finally, we have presented an unsupervised algorithm for finding the upright orien-tation for man-made models based on low-rank matrix theorem. Most man-made models can be posed at a unique upright orientation which is consistent to human sense and its intrinsic property. We have found that projections of the models could be regarded as low-rank matrices when they have been posed at axis-aligned orientations. Thus, we it-eratively rotate the models by using the recently presented TILT algorithm, in order to ensare that their axis-aligned projections have optimal low-rank observation. After that, the model will be aligned with the axes, the upright orientation can be picked up from the six candidate axis-aligned orientations by using simple geometric analysis. The proposed algorithm is able to be performed independently without additional training data of mod-els. Experimental results have been made to clarify the effectiveness and intuitiveness of the algorithm.