Research On The Key Technology Of High Efficiency Video Coding

Recently, with the rapid development of network communication and multimedia technologies, video applications have been widely used in the human daily life. The ubiquitous video applications have produced huge amounts of video data. Furthermore,the video contents are presented to be High Definition(HD) and Ultra HD. Compared with the standard definition videos, the resolutions of the HD videos are larger and the visual quality is better. However, the uncompressed video data are so large that they need to be compressed for being employed in the actual storage and transport. Thus,the video compression is a key technique of the video applications. Meanwhile, high requirements are put forward on the video coding technologies. To meet the requirements,ITU-T and ISO/IEC jointly developed a new generation of video coding standard, called High Efficiency Video Coding(HEVC). Compared with the previous standards, HEVC has the higher compression efficiency and the better image quality. Thus, the HEVC standard has bright prospects. However, the efficient coding performance also brought huge coding complexity, which severely affecting the application and popularization of the standard. Therefore, reducing the coding computational complexity effectively is a successful factor of the HEVC standard, and the urgent need of the HD applications.Since the process of encoding a Coding Unit(CU) includes almost all the modules in the HEVC, the CU decision occupies most of coding time in HEVC. Thus, to reduce the coding complexity, this thesis mainly studied on the fast CU decision algorithms of the intra frames and inter frames. The contributions include as follows:Firstly, a non-normalized gradient histogram(n-HOG) codebook based fast CU decision algorithm is proposed for HEVC intra prediction. The relationship between the strongest correlation and the rate-distortion cost is analyzed based on the image correlation model. The CU split flag is proved to be closely related to the n-HOG of the CU.Thus, the n-HOG feature is defined to predict the CU split flag. A fuzzy clustering based method is designed to train the codebook, which is used for the CU decision. According to the characteristics of the intra prediction, the method has provided an efficient feature,which can effectively reduce 30% of the total coding time.Secondly, a template matching based fast intra CU decision algorithm is proposed with the n-HOG feature. A first-in first-out queue is designed to store the CU coded information. The nearest neighbor of the similar CUs is searched to help determining the CU split flag. This algorithm is inspired by the classic method in the field of image processing, called template matching. It can effectively use the information of the encoded CU. The encoded information has important reference value for the CU decision.Thirdly, a pyramid motion divergence(PMD) based fast CU decision algorithm is proposed for the inter prediction. The characteristics of the pixel movement between the adjacent video frames are analyzed. Based on the image correlation model, the ratedistortion cost is proved to be positive correlation with the variance of the motion vectors,which is defined as motion divergence. Then, the PMD feature and the down-sample optical estimation based PMD extraction method is proposed for the CU decision. Finally,the nearest neighbor searching based fast CU decision algorithm is proposed. According to the characteristics of the inter prediction, the method has provided the PMD feature, which can effectively reduce the computational complexity inter CU decision.Fourthly, a markov random field(MRF) based fast inter CU decision algorithm is proposed. The CU decision is simplified as a labeling problem. The CU split flags are defined as an MRF. Then, the classic graph cut theory is used to solve the CU split flag.Based on the movement characteristics of the pixel, the variance of the absolute difference is proved to be positive correlation with the rate-distortion cost. Finally, a rate-distortion cost based maximum a posteriori method is proposed to determine whether an Unsplit CU should be further split or not. The MRF model can effectively combine the CU features with the context information effectively.Finally, a two-layer motion estimation based fast CU decision is proposed for inter prediction. First, a two-layer motion estimation method is designed to search the SAD cost of the current CU and its sub-CUs. Then, the concept of the motion compensation cost is proposed. The exponent model of the SAD cost and the quantization parameter is defined to express the motion compensation cost. A threshold is derived from the express model to determine the CU split flag. This method employs the latent SAD costs of the sub-CUs in the fast CU decision and reduces the computational complexity, effectively.In addition, it has provided a feasible solution for implementing the fast CU decision in the parallel framework and on the hardware development.