Noncausal image prediction and reconstruction

Abstract

Summary form only given. Prediction of the value of the pixels in an image is often used in image compression. The residual image, the difference between the image and its predicted value, can usually be coded with fewer bits than the original image. In linear prediction the value of each pixel of an image is estimated from the value of surrounding pixels using a predictor P. In noncausal prediction pixels surrounding the pixel to be predicted are used. In causal prediction only "earlier" pixels are used. Usually noncausal prediction offers better prediction than causal prediction because all pixels surrounding the pixel to be predicted are considered. The reconstruction of the image from the residual after noncausal prediction is more difficult than when causal prediction is used. This paper explores two methods of reconstruction for noncausal prediction: iterative reconstruction and direct reconstruction. As an example, the effect of quantization of the residual on the reconstructed image is considered. It shows an improved image quality using the noncausal predictor.