| We address two important problems in multimedia signal processing: error resilient video transcoding and image registration. Video communication through wireless channels is a challenging problem due to the limitations in bandwidth and the presence of channel errors. Since many stored video sources are originally coded with high bit-rate, they cannot be sent directly through the wireless channel. Therefore, transcoding is needed to reduce the rate of the video source. However, given the existence of channel errors that can easily corrupt the quality of video, there is also the need to make the bitstream more resilient to these transmission errors. Image registration can be applied to video mosaicing, medical imaging, camera calibration, and radar image analysis. We propose registration techniques that can achieve sub-pixel accuracy.; In the first part, we address the problem of video transcoding for transmission in an error-prone environment. One key issue is to balance the rate used for the video source with bits allocated for error resilience such that the end-to-end distortion is minimized under the given rate constraint and channel condition. The rate for the video source is reduced by requantization, while the bits for error-resilience are controlled by the insertion of resynchronization markers and intra-coded blocks. We establish accurate rate-distortion (R-D) models for video source requantization that account for inter-frame dependencies, as well as R-D models for error propagation in motion compensated video. Based on these models, we propose an optimal bit allocation scheme. In order to account for the inter-frame dependencies, this allocation scheme operates on a Group-of-Pictures (GOP). The proposed optimal scheme achieves better PSNR performance than the fixed heuristic bit allocation scheme. We also propose a sub-optimal scheme that is suitable for a real-time implementation. This suboptimal scheme significantly reduces the complexity of R-D model estimation while keeping the loss in performance to a minimum.; In the second part, we present methods to accurately align images related by affine or projective transformation. We achieve this by matching and registering plane curves in the related images. We first present algorithms for aligning curves modelled by affine transformation. Superior accuracy and efficiency is achieved by the use of super-curve and B-spline fusion. (Abstract shortened by UMI.)... |
