Rate/distortion-based combination of multiple DCT transforms for video coding

This paper proposes some early results concerning the combination of discrete cosine transforms (DCTs) of multiple sizes in a video encoder. The main difficulty consists in finding a way to obtain a fair competition between several transforms. For that purpose, we derive that the quantization of all transformed coefficients regardless of the DCT size with the same uniform linear quantizer allows to obtain an equivalent level of distortion among all transforms. The selection of the best transform on a particular image area is then made by a rate/distortion framework, which jointly optimizes motion and residual signal coding. Systematic objective and subjective improvements are observed when compared to using a single DCT size.