Two-pass MPEG-2 variable-bit-rate encoding

Many MPEG-2 encoding applications are realtime; this implies that the video signal must be encoded with no significant lookahead. However, there exist non-real-time applications that do enable us to first analyze a video sequ ence entirely, and, using the analysis results, to optimize a second encodi ng pass of the same data. One example of such an application is the digital video disk (DVD), which is designed to facilitate a variable-bit-rate (VBR ) output stream. In that case, it is possible to let the MPEG-2 encoder pro duce a video sequence with a constant visual quality over time. This is in contrast to constant-bit-rate (CBR) systems, where the rate is constant but the visual quality Varies with the coding difficulty. This paper describes a two-pass encoding system that has as its objective to produce an optimiz ed VER data stream in a second pass. In a first pass, the video sequence is encoded with CBR, while statistics concerning coding complexity are gather ed. Next, the first-pass data is processed to prepare the control parameter s for the second pass, which performs the actual VER compression. In this o ff-line processing stage, we determine the target number of bits for each p icture in the sequence, such that we realize the VER objective. This means that the available bits are appropriately distributed over the different Vi deo segments such that constant visual quality is obtained. To be able to q uantify the constant visual quality, perceptual experiments are described a nd a practical model is fitted to them. Exceptional cases such as scene cha nges and fades are detected and dealt with appropriately. We also ensure th at the second-pass compression process does not violate the decoder buffer boundaries. Finally, the encoding is performed again, but now under control of the processed first-pass data. During the running of this second pass, a run-time bit-production control mechanism monitors the accuracy and valid ity of the first-pass data, correcting errors in prediction and observing t he buffer boundaries. Results are compared to CBR operation.