Study On Video Compression Based On Elastic Motion Model

Nowadays, with the fast popularization of high-definition, ultra-high-definition video and3D movies, video applications are developing toward high-resolution and high frame rates. In the case of limited transmission bandwidth and storage space, efficient compression of video information has become the key technology. Under such a background of new applications, High Efficient Video Coding (HEVC), the newest generation video coding standard, has pushed video standardization process to a new height, which means HEVC is able to using approximately30-50%less bit rate on average compared with the previous generation video standard H.264/AVC at the equivalent subjective quality.However, HEVC still uses the previous hybrid coding framework. Especially in the inter prediction, Block Matching Algorithm (BMA) based on a rigid translational motion model is still used in HEVC, and BMA has been already adopted since the first generation video coding standard.The conventional BMA has the following three limitations:First, constant motion vector fields, where all the pixels in a block are assigned the same motion vector; second, translational motion model is not able to capture camera zoom, rotation and complex motions such as the deformation of objects; third, complex motion in blocks is usually approximated using BMA motion model with smaller blocks, which will bring more coding bits of motion vectors. From a rate-distortion optimization point of view, these limitations of conventional BMA directly impede the performance of video compression to be better. So, study on effective non-translational motion model is one of the important issues to be faced for video development in the future.In this paper, we study on an elastic motion model based on2-D discrete cosine basis functions, and verify its performance in the inter motion prediction in comparison with conventional block translational motion model.2-D discrete cosine basis functions based elastic motion model is able to effectively match elastic motion vector fields such as camera zoom, rotation and deformation with less motion parameters. To estimation elastic motion model parameters, an iterative Gauss-Newton gradient descent nonlinear optimization algorithm is used. Finally, we integrate the inter prediction module based on elastic motion model and apply it to the coding model HM12.0in HEVC. Experimental results including subjective and objective qualities show that, compared with HEVC, our improved HEVC based on elastic motion model achieves on average3-12%bit rate reduction with approximately1%average PSNR distortion at medium or low bit rates, while subjective quality is almost the same.