Low Bit Rate Video Coding with Geometric Transform
چکیده
This paper describes how the combination of geometric transformations in both the spatial and pixel domains can be used for video coding. The geometric transformation applied in the spatial domain is able to track movements such that, for low bit rate video coding, the motion vectors alone can be used to represent a complex motion of moving objects. The introduction of such a transformation in the pixel domain allows the recovering from non-uniform intensity changes, as well as uncovered background. The presented method achieves a compression eeciency far superior than that obtained by using conventional block matching motion compensation. In order to reduce the computational complexity, image segmentation, fast transformation and interpolation algorithms are used. During the past decade a signiicant progress in the eld of video coding has been made, and various intra-frame coding techniques have been devised. However, most of the eeciency of low bit rate video coding is due to the motion compensation. In this sense, half-a-pixel accuracy block-matching motion estimation has been used in MPEG 1], in order to improve the coding eeciency of this codec over H.261 2]. In the mean time H.263 3], a successor to H.261 for lower bit rate applications, uses also half-a-pixel accuracy. It is important to note that at low bit rates, any overhead, like those of the motion vectors can be very costly. Despite this, H.263 uses it and still is more eecient than H.261. Motion compensation eeciency can be further improved by using geometric transformations to deform the moving quadrangular patches in order to achieve sub-pixel accuracy. In our previous work 4, 5], we have shown that this technique is so eecient that the motion compensated predicted (MCP) error for most parts of the picture is not needed to be coded at all. The main reason for using spatial transform motion compensation instead of the conventional block matching for low bit rates can be given as follows. In the conventional block matching motion estimation (BMA), the motion of objects is assumed to be purely translational. This means that all the pixels within a moving area have the same speed. To reduce the error of this assumption, small blocks are used, in order to reduce the disparity of motion of the pixels inside the block. This causes however an increased motion vector information overhead. In the sense that small blocks reduce the coding error, an ideal motion compensation would …