Adaptive Down-Sampling and Super-Resolution for Additional Video Compression

While almost all down-sampling based video codecs gain additional compression at the expense of image degradation, we set a good example of achieving both large compression and even better reconstruction quality. Such progress is realized by: (i) minimizing the introduction of information loss with a proposed decomposition-based adaptive down-sampling method so that more reserved pixels can be allocated to image details where human visual perception is more sensitive. Specifically, a modified content complexity measurement is put forward and the optimum down-sampling rate is adaptively selected with a customized formula; (ii) maximizing the information compensation via a content-adaptive super-resolution algorithm, which is accelerated and optimized by two stages of pruning to select the closest correlated dictionary pairs. Extensive experiments support that, by using prevailing H.264 codec as benchmark, the proposed scheme achieves 5 times more of additional compression and the reconstruction quality outperforms other state-of-the-art approaches, and even better than decoded non-shrunken frames in human visual perception.