A novel approach to error recovery in the transmission of transform coded pictures

In this paper, a mechanism for the recovery of lost data in the transmission of JPEG coded images over packet networks is presented. The loss of a packet conveying coded block data brings to the unsuccessful reconstruction of the relevant picture area. The proposed method allows to recover a subset of DCT coefficients by exploiting the structure of the correctly received neighboring blocks, in the very critical case when the loss concerns all the block coefficients. For this purpose a neural predictor was specifically designed and suitably trained with a set of carefully studied synthetic patterns. It directly works in the transformed domain, allowing a real time reconstruction of damaged blocks. The results achieved show that the method has very good reconstruction capability both on synthetic and natural test data