Exploiting the directional features in MPEG-2 for H.264 intra transcoding

This paper presents an MPEG-2 to H.264 intra frame video transcoder. The DCT coefficients gathered from the MPEG-2 process are used to estimate the directional features in a picture. The estimated directional features are then used to compute the intra prediction modes in the H.264 encoding stage. The INTRA 16×16 vs. INTRA4×4 MB coding mode decision is made based on the variance of the DC coefficients of the 8×8 DCT blocks in the input MPEG-2 video. The proposed approach completely avoids the MB mode and prediction mode computations in the H.264 encoding stage, there by reducing the transcoder complexity substantially. The results show that our approach results in over 30% drop in the intra frame encoding time with a negligible drop in PSNR. A comparison with a cascaded reference transcoder is presented.