Research On Semantic-Based Coding Of Portrait Sequences

Semantic-based coding is an active topic in the field of multimedia information processing, and it is involved with the theories and technologies of computer vision, computer graphic and mathematical statistics etc, so this dissertation is a subject with challenge in theory and difficult in realization. The portrait sequence is a commonly used form in video. Consequently, this dissertation focuses on some key issues of semantic-based coding of portrait sequences. And the study is carried out in the following aspects.1. Method of precisely extracting object contour is investigated. Precisely extracting contour is the base of video representation. In most cases, the video object is noised or has the extreme concave edge, which often makes the contour extracting inaccurate. The active contour model, a famous dynamic algorithm, is commonly used to extract the object contour,and GVF Snakes is one of its improved algorithms. Though it overcomes the shortcoming of the traditional Snakes, GVF Snakes is sensitive to the extreme concave contour and noise that often cause some control points of curve converge to the positions of the local minimum value some times. Genetic Algorithm (GA) can optimize function that does not satisfy the continuity or differentiability, so it is an effective optimization algorithm. At the same time, the search process can be sped up by parallel computation on the sub-population fitness function and the new individual generation in Parallel Genetic Algorithms (PGA). The robust method, which improves the GVF and optimizes search by PGA, is proposed to extract contour veraciously in this section. Due to the information exchange between adjoining points by chromosome crossover, the points are driven to optimization positions in the fined-grained model. Experimentation shows that this method can extract the extreme concave and noise-disturbed contour accurately and robustly.2. Algorithm of the semantic object extraction is studied. AS an ill-posed problem, segmentation should resort to restriction to obtain a reasonable result. In order to obtain a semantic object, the traditional segmentation is restricted with the prior knowledge such as the shape or structure in semantic segmentation. The head and shoulders are generally regarded as the same video object plain (VOP) in the head-shoulder sequences, a commonly used form in video. The head-shoulder contours from different views are separated into five classes according to theshape characters, and shape model is defined respectively. To effectively remove redundancy and reduce the description complexity, the principal component analysis (PCA) is used to obtain shape characteristics by training samples in each class. Maximum likelihood criterion is used to classify new contour sample, and the shape template is synthesized with corresponding principal variables. In order to match head-shoulder contour correctly, an algorithm is proposed to control and restrict the deformation of discrete deformable template by multi-restrictions: edge, motion and curvature. Due to the lack of motion information, the segmentation becomes more difficult in static news images. Finally, a new means is designed to obtain semantic object based on the color and shape model.3. Motion estimation based on wire-frame model is researched. Firstly,two commonly used algorithms ( predictive least squares(PLS) and extended Kalman filter) are introduced, and the reasons are also analyzed that they are sensitive to noise and easy to diverge. An improved EKF method is presented by self-adaptive parameter-amend technology and smooth filter, which not only overcomes disadvantages of the computational oscillation and divergence, but also estimates long-term motion accurately and robustly. Experimental results indicate the efficiency of this improvement. To avoid heavy computation caused by matrix inversions and ensure its convergence, relaxation-iterative search algorithm( RSA ) is utilized to estimate motion. RSA not only estimate accurately but also save memory, so it is more suitable for SoC.4. Hybrid video coding system is designed. Although SPIHT compress static image effectively, it often loses details in very low bit-rate for video, which should blur restoration image. Due to the difficulty in modeling general video scene, the extensive use of model-based coding is limited despite the satisfying subjective quality of synthesize image. In order to operate semantic object easily, a model-assisted hybrid coding system is presented to integrate the two above-mentioned coding methods. Some key techniques are depicted: the frame-ratio control, texture compressing method, clip-and-paste facial expression synthesis and image synthesis. Performance of system is shown by the result of simulation experiment.