Single-pass constant- and variable-bit-rate MPEG-2 video compression

IBM Journal of Research and Development, Jul 1999 by Mohsenian, N, Rajagopalan, R, Gonzales, C A

and bit usage of all encoders for the LaBk sequence. A comparison of the average PSNR numbers for the pictures in the special-effect region is given in Table 3.

Also presented in Table 3 are results for the Sprite sequence, which is composed of a segment where a fade to (from) black from (to) natural video material takes place. Fades are a special class of temporal image discontinuities which can seriously challenge the robustness of the motion-estimation technique of the encoder. As a result, 13- and P-picture types do not do well during fades, and it is better to pre-analyze a collection of pictures, prior to encoding, and label them all-I if a fade is detected. Our single-pass encoders do not require a pre-analyzing step and are capable of adjusting their RC parameters in real time to compensate for the luminance changes of the pictures within the fade. Compression results for the Sprite sequence are given in Table 2. Finally, as the last test sequence, we encoded Singer, which comprises flashes of different intensities. For this case a small improvement in the PSNR performance is obtained, as indicated by the tabulated results of Table 2. against a predefined threshold. A reference picture can be either I- or P-type. Since a GOP always starts with an I-picture, and we have assumed a fixed GOP structure throughout this paper, intensity differences are computed between a current I-picture and a previous P-picture. Let Y,11 represent the average intensity of the luminance component of an I-picture (in temporal position n) to

be encoded at the start of the GOP, and let ... represent the average intensity of the luminance component of a previous P-picture just encoded. The The set of weights {Omega^sub 1^, Omega^sub 2^, Omega^sub 3^} is chosen so as to give importance to certain statistical measures. We label the VBR encoding scheme which adopts the scene-change criterion of Equation (46) as VBR w/GLSCD. The impact of the scene-change detection strategy on the quality

of the VBR streams is assessed by compressing video sequences mpeg, mixb, and GreySep. A comparison of the PSNR performance of the VBR encoders with and without the GLSCD formulation is provided in Table 4. The importance of the GLSCD addition is not reflected by the PSNR improvements of Table 4. An examination of the PSNR figures (not shown here) revealed that detecting video discontinuities and applying the median filtering of (40) and (41) cleans up some of the undesired bit targets that are allocated between the scenes of different complexities.

We also simulated the highly detailed video sequences mpeg and mixb using the GOP bit-irregularity detection (GBID) methodology defined by (42). This approach improves the quality of the difficult segments of sources such as "Cheerleaders" and "Mobile and Calendar" of the mpeg sequence and "Skyscrapers" of the mixb sequence, as shown by the PSNR plots of Figures 13 and 14. In these regions, GOP bit irregularities are filtered, and a more intelligent bit distribution is achieved. Table 5 summarizes average-sequence PSNRs obtained by using the VBR encoders with and without incorporating the GLSCD and GBID methods.