Abstract: In many codecs a predicted image is calculated by motion compensation on a block-basis. One motion vector for each block is determined and the prediction used is a displaced block from the previous image. This method has one major disadvantage: blocking artifacts. In this text we investigate the use of an interpolated motion field to solve this problem. A discrete number of motion vectors is used to characterize the motion field. With these vectors a motion vector for each pixel is calculated. If we use this motion field for prediction all the blocking artifacts disappear thus giving a visually better prediction. Using the motion vectors resulting from the block-based codec, which are not optimized for this interpolation, and interpolating them has one drawback. It decreases (compared to blockmatching) the signal to noise ratio of the prediction, leaving a bigger error to code, and thus the gain of a visually good prediction is partially lost. We propose an algorithm for the calculation of motion vectors that solves this problem. Knowing the fact that we will interpolate, we are able to guess the motion field on a pixel basis for each possible motion vector. Thus we are able to get an optimal motion vector for this guessed pixel based motion field which is very close to the resulting pixel based motion field. By interpolating these motion vectors we are able to increase both the subjective and objective quality of the predicted image. !3
展开▼
