Research On Complexity Scalable Video Coding

With the development of mobile and pervasive computing technology, videocommunication applications, such as video telephone, video conferencing and videosurveillance, are expanding across various terminals. Herein, video communicationfaces a big challenge from matching between high complexity of video coding anddifferent computation capacity of various terminals, so it is necessary to enablescalable complexity of video coding. Complexity Scalable Video Coding (CSVC)focuses on studying video coding algorithms and framework, which can dynamicallyadjust computation load of video coding. As a scientific issue, CSVC belongs to thevideo coding optimization field. It needs study the fundamental problems of videocoding such as the spatial and temporal complexity analysis andcomplexity-rate-distortion model. From the viewpoint of applications, CSVC willprovide video communication across various terminals with key technologysupporting.The complexity scalable algorithms of both motion estimation and DCT foradvanced video coding technology are studied in this work, so as to provide adaptivecoding framework to various terminals. The main contribution is as follows:1. Fast mode decision algorithmTwo techniques, multiple frames referencing and multiple modes coding lead tothe complexity increasing rapidly. Considering the two factors comprehensively, a fastmode decision algorithm for inter frame coding is proposed. By adaptively selectingcandidate modes to eliminate the unnecessary modes, this algorithm can find the bestmode quickly. Simulation results show that average complexity reduction exceeds85% and average quality degradation is only about 0.07dB compared with full searchwhen using 5 reference frames. By adapting the number of coding modes, thisalgorithm can be used to control the motion estimation complexity with multiplecontrollers.2. Search center adaptive motion estimation algorithmSub-pel motion vector search leads to the coding complexity increasing quickly.The existing fast motion estimation algorithms conduct integer, half and quarter pelsearch sequentially, and the predicted motion vector information has not been madegood use to reduce search points. Based on analyzing the distribution of motion vectordifference, a search center adaptive motion estimation algorithm is proposed in thiswork. The search center is determined by comparing the coding costs, which arecomputed by using zero motion vector, predicted motion vector (PMV) and thenearest integer motion vector from PMV respectively. Simulation results show thatthe integer pel search process of 25% blocks can be skipped for the benchmarksequences with complex motion. In addition, the complexity will decrease furtheronce the algorithm is applied together with the other fast algorithms based onreducing search points. By adapting the search center and search range, this algorithmcan be used to control the motion estimation complexity with multiple controllers.3. Complexity scalable integer DCT algorithmInteger DCT accounts for a big proportion of real time video coding computation.A novel model to identify the type of pruned DCT block is built for both luma andchroma component, and based on the model we presents a complexity scalable DCTalgorithm. Simulation results show that compared with classical Laplacian model andthe model based on coefficient scaling, the DCT algorithm based on the model canreduce above 20% DCT complexity with negligible performance loss, thus DCTcomplexity can be controlled within the target complexity.4. Complexity constrained rate distortion optimization modelThe rate distortion optimization with complexity constrained is a problem tooptimally selecting encoding parameters under specified complexity. In this work,quantitative analysis on the complexity of video coding is conducted. Furthermore, acomplexity constrained rate distortion model is built based on the analysis of theinfluence on coding performance of different encoding parameter settings. With thehelp of the model, a scheme to select optimal encoding parameters is designed, whichcan improve coding performance as greatly as possible with specified computationconstrain.