Improved machine learning techniques for low complexity MPEG-2 to H.264 transcoding using optimized codecs

This paper discusses techniques for efficiently implementing a Mpeg-2 to H.264 video transcoder. The transcoding results reported in the literature are based on a reference implementation and may not reflect the true performance gains obtained in real world systems. We have developed low complexity transcoding algorithms and have implemented these solutions using highly optimized encoder and decoder implementations available from Intel. The transcoding algorithms are based on exploiting the mode decision knowledge inherent in the decoded MPEG-2 data. Machine learning techniques are used to make accurate and low-complexity H.264 MB encoding mode decisions. The results show that the proposed transcoder reduces the complexity by 50% without a significant loss in PSNR. This performance improvement in production quality transcoders, and demonstrates the practicality of machine learning based video transcoding algorithms.