| Both texture synthesis and video temporal segmentation are research interests in computer graphics, image processing and computer vision.Traditional texture mapping technique can not be widely used, because it causes texture seams and distortions when rendering large area textures or surface textures. Texture synthesis is suggested to overcome the flaw. Now, texture synthesis can be divided into procedure texture synthesis (PTS) and texture synthesis from samples (TSFS). By using PTS, highly realistic textures are directly generated on surface through simulating the physical generation procedures of objects. But it is too tedious to test the parameters again and again for each type of texture in order to get appropriate ones, and even sometimes satisfactory results can not be obtained. The basic idea of TSFS is that the continuity and similarity of textures'structure are maintained when large area surface textures are generated from given small samples based on their texture characteristics. This technique draws more and more attention, because it does not only overcome flaws of texture mapping, but also avoids the tedious procedure to test parameters of PTS. TSFS can be divided into three catalogs, which are 2d texture synthesis, texture synthesis on surface and video texture synthesis. The existing 2d texture synthesis methods mainly contain pixel-based methods and patch-based methods. The principle of pixel-based methods determines their low speeds to synthesize textures. Some scholar proposed improved algorithms to speed up pixel-based methods, but these algorithms are too complicated and too hard to implement. Speeds of patch-based methods are quick enough, but their principle determines that textures with non-rectangular shapes can not be fully sampled and mosaics are generated from textures with narrow shapes. Video texture synthesis algorithm based on transitions is a typical video texture synthesis method from samples. To synthesize video texture, this algorithm needs implement several complicated steps, such as calculating differences between frames, calculating matrix of transition costs, avoiding dead-ends, pruning part of transitions, calculating future costs and creating video loops etc.; The preprocessing is also complicated and the computation costs a lot before creating video sequences; Dynamic Programming used to create video sequence allows only backward transitions, which reduces the multiplicity of synthesized results.Video temporal segmentation is basis of the whole video content analysis. Only when video sequences were divided into shots, the other steps could be efficiently implemented, such as key frame extraction, video thumbnails and video sequence recognition etc. Therefore, video temporal segmentation is a very active research direction with the longest research history and the most published papers in video analysis. The critical part of video temporal segmentation is shot boundary detection, which includes detecting shot boundaries i.e. to find the first frames and the last frames of shots and recognizing types of shot boundaries to be cut transitions or gradual transitions. Now methods to detect cut are satisfying, but methods to detect gradual are not satisfying yet. Moreover, most methods to detect not only cut but also gradual need search cuts and graduals in different ways or steps. For shot boundary detection methods, the most important factors that influence results of detection are missed detections and false alarms. Missed detections mean that actual shot boundaries are missed due to poor similarity scale and measure principle. False alarms mean that continuous frames are treated as shot boundaries, which are caused by camera/object motions and abrupt changes of luminance. Therefore, quality of shot boundary detection algorithm can be improved by reducing missed detections and false alarms.In this paper, several problems concerning the two directions mentioned above are researched, such as 2d texture synthesis, video texture synthesis, shot boundary detection algorithm and its application in video watermark area. Furthermore, a 2d pixel-based texture synthesis method using particle swarm optimization (PSO), a video texture synthesis method based on genetic algorithm (GA), a shot boundary detection algorithm based on PSO classifier and a shot boundary detection algorithm based on directional empirical mode decomposition (DEMD) are proposed.Pixel-based texture synthesis method using PSO is an eclectic scheme between pixel based texture synthesis methods and patch-based texture synthesis methods. It changes the mode that search throughout of traditional pixel-based methods. It looks for matched pixels from samples and pastes them to result textures by using PSO. PSO is developed from simulation of simplified social models. In PSO algorithm, each individual always updates itself according to the current best solution it has traveled and the current global best solution, and then the approximate best solution for problems would be found quickly. Pixel-based texture synthesis method using PSO has quickened the traditional pixel-based method. Although it is a little slower than patch-based methods, it overcomes the flaws of them that textures with non-rectangular shapes can not be fully sampled. The number of particles and the maximum number of iterations that influence the efficiency of algorithm are discussed. Now, the two parameters need to be set by humans. In future research work, it is expected to find the relation between property of image and the parameters, and then they could be automatically set when synthesizing various images.In the video texture synthesis method based on GA, video texture synthesis is translated into a combinatorial optimum problem of frames. GA is introduced into the processes of synthesizing video textures and the new video texture synthesis method can be used to generate continuous and infinite streams of video from a finite video clip. In this method, frames are regarded as genes, video sequences are regarded as chromosomes and population is composed of chromosomes. Chromosomes are continuously changed by selection operator, crossover operator and mutation operator of GA and evolved to satisfying solutions i.e. continuous video textures in vision. Instead of more complex pre-processing of source video clip, only an appropriate fitness function is needed in this algorithm. Comparing with representative video texture synthesis algorithms nowadays, this algorithm has less computational complexity and improves the efficiency of synthesis. In the video texture synthesis method based on GA, the continuity of video sequences is paid much more attention than the structure of them. A more appropriate fitness function for both continuity and structure is expected to be found in future research work.Furthermore, several similarity scales and measure principles are experimented and compared in this paper, and a more appropriate one is selected to calculate differences of frames. A segment genetic algorithm (SGA) is also used to synthesize video texture in the improved algorithm. Comparing with the method mentioned above, the improved one has quicker speed and better quality to synthesize video texture.In shot boundary detection algorithm based on PSO-Classifier, difference curves of U-component histograms are used as the characteristics of the differences between video frames in order to make the algorithm more sensitive to gradual transitions and not sensitive to camera motions and object motions. After filtering difference curves by a Slide-Window Mean Filter method, there are figures like rectangles at the locations of cut transitions and figures like triangles at the locations of gradual transitions. In order to recognize these figures with remarkable characteristics, cut training set and gradual training set are established. A Nearest Neighbor algorithm applying PSO is used to classify the figures to be detected. Types of shot boundaries are judged according to the characteristics of their figures. Finally, shot boundaries are detected and recognized. Experimental results show that the performance of this algorithm is comparable with the existing shot boundary detection algorithms'. Now, only cut training set and gradual training set are established, and the more types of graduals need be recognized, the more training sets and matching rules should be established.Shot boundary detection algorithm based on directional empirical mode decomposition (DEMD) is a detection algorithm with high precision to cut transitions. Abrupt change of luminance is one cause of false alarms. In our experiments, after decomposing two similar frames with much different luminance in a video sequence by DEMD, the difference between the intrinsic mode functions (IMF) with the lowest frequency of the two frames is the most remarkable. In order to reduce false alarms caused by abrupt change of luminance, the frames are gotten rid of their IMFs with the lowest frequency and used to calculate differences between frames. Object/camera motion is another cause of false alarms. A block tracing motion compensation method is applied to avoid false alarms in this paper. In the method, a reference frame C and a remember matrix R are established. C and R are updated through comparing differences between corresponding blocks of adjacent frames. Elements in R are used to judge whether there are shot boundaries in frames or not. Shot boundary detection algorithm based on DEMD is no longer sensitive to abrupt change of luminance, object motion in shots, motion and zoom in/out of cameras, so that the precision of detecting shot boundaries is improved. Using independent component analysis (ICA) to analyze videos segmented by the method mentioned above, a series of independent component frames is obtained, then applying an improved wavelet-domain quantization-based image watermark algorithm to embed watermarks into these frames to realize the embedding of video watermarks. Experimental results show that video watermarking method in this paper is robust enough against several kinds of watermark attacks; it is more robust against attacks such as frames discarding and decreasing in a same time than traditional video segmentation methods based on histogram.
|
PDF Full Text Request