Comparison of image transforms in the coding of the displaced frame difference for block-based motion compensation

This paper compares the performance of various image transforms in the coding of residual images obtained after applying block-based motion compensation to a video sequence. The results show that the performance penalty associated with using a “sub-optimal” transform (ie: one other than the KLT) is generally larger in the case of residual images than it is when still images are coded. However, the results also showed that the KLT is not the optimal transform to use at all data rates. In addition, the DCT does not work significantly better than the simpler Haar, Walsh-Hadamard and Slant transforms when operating on the residual images