Post-processing of DCT-based image and video decoded data in error-prone environments

• Main Author: Shirani, Shahram
• Language: English
• Published: 2009
• Online Access: http://hdl.handle.net/2429/10913

Compressed image and video bit streams are very sensitive to channel errors and may be altered or lost during transmission. Error concealment by post-processing intends to reconstruct lost visual information by exploiting the correlation between the image/video data. The applications of concealment of errors in coded visual information include visual communication over unreliable channels such as wireless networks and the Internet. For most types of encoders and input data, coded visual information consists of a collection of coded texture (DCT coefficients), shape and motion information. In this thesis we present concealment methods for errors in texture and shape information, and address the concealment of errors in motion data in conjunction with its corresponding texture or shape information. The method developed for concealment of errors in coded texture involves compensation of the effects of the missing data on the rest of the texture information and then using, a deterministic or a statistical algorithm for the restoration of missing information. The deterministic algorithm achieves a good performance level in the reconstruction of edges. The statistical algorithm which is based on maximum a posteriori (MAP) estimation, employs an adaptive Markov random field (MRF) as the image a-priori model. The adaptation enables the estimation procedure to incorporate more information without a dramatic increase in computational complexity. MAP estimation is also employed for the reconstruction of missing shape data. Although it uses an adaptive MRF, the estimator is different in the sense that it is designed for binary shape information. In the second part of the thesis, we evaluate the performance of the developed concealment methods for three different types of coded visual data: baseline JPEG coded still images, H.263 coded video and MPEG-4 coded video. Our experimental results demonstrate that the methods presented in this thesis achieve consistently good computation-performance tradeoffs, making them very beneficial for real time communication over error prone networks. In fact, the proposed error concealment methods can lead to acceptable visual quality at loss rates as high as 20%.