Quality and Motion Estimation for Image Sequence Coding

Since the advent of television obtaining high perceived quality using a limited bandwidth has been an important issue. This thesis proposes methods inspired and based on the human visual system for exploitation of temporal and perceptual redundancy in image sequences to achieve lower coding rate and/or higher visual quality. Particularly relevant in the present context are the indications that the visual cortex contains cells that are selective in orientation and frequency but invariant to the phase of the stimuli. For this reason and for computational efficiency, a spatial quadrature filter bank is generated initially using combinations of simple sequential 1D filter kernels. The temporal redundancy is reduced using motion compensated prediction. Motion compensation predicts the frame to be transmitted by moving the previous decoded frame according to its motion, so called backward coding. This means that no information about motion needs to be transmitted to the decoder. The irregular sampling of the prediction, due to motion estimates with sub-pixel accuracy, is dealt with using a new method, continuous normalized convolution. Image values on an interlaced sampling grid are estimated from the irregularly sampled predictions using integration of signals and certainties over a neighbourhood employing a local model of both the signal and the integration filter. The motion is estimated from products of quadrature filter bank responses, generating phasebased channels tuned for velocity. This new approach includes learning parameters for motion estimation and introduces multiple hierarchical motion estimation to achieve estimates with high spatial resolution. Using the quadrature filter bank a new quality estimator, based on a human visual system model, is generated. The quality estimator can be used to reduce the perceptual redundancy of non-visible errors and is intended to improve coding performance in such regions.