Efficient Video Compression

While conventional codecs “undo” the operation performed by the encoder, we consider the decoder to be solving an estimation problem where the best estimate of the source sequence is generated based on an optimality criterion, the available bistream, and prior knowledge about the source sequence.

Our exploration of this paradigm started with encoding in an intermediate resolution. In this framework, a high-resolution (HR) sequence is downsampled prior to the compression. On the decoder side, original resolution is recovered by super-resolution techniques. In order to incorporate content information into the system, segmentation is carried out, based on PCA and motion analysis on a group of pictures (GOP), each frame is partition into areas of motion (M), no motion and flat (F) and texture (T). The resulting partitioning information, called Motion-Texture (TM) map, is used to drive downsampling process operation in the encoder and super-resolution and interpolation operations in the decoder, by transmitting it as auxiliary information. In this approach, M type blocks are downsampled and reconstructed using a super-resolution technique. The areas that are flat are downsampled and reconstructed using interpolation and textured regions are downsampled according to the block prediction type.