Techniques for Improving the Error Resilience of Mpeg-2 Codecs

This paper is concerned with the performance of MPEG-2 compressed video when transmitted over noisy channels, a subject of relevance to digital terrestrial television. We present the results of introducing errors into MPEG-2 video, and propose techniques for substantially improving the resilience of MPEG-2 to transmission errors without the addition of any extra redundancy into the bitstream. We find that it is errors in variable length data which cause the greatest artefacts as errors in these data can cause loss of bitstream synchronisation. We achieve resynchronisation using a technique known as error-resilient entropy coding (EREC). Finally we improve the error-resilience of differential coded information by replacing the standard 1D-DPCM with a more resilient hierarchical pyramid predictor. INTRODUCTION Figure 1 shows the first MPEG-2 (1) intra picture of a coded sequence. It can be clearly seen that a few errors can cause a cause large areas of the picture to be corrupted. Fig. 1. An MPEG-2 Intra Picture. BER = 0.01% It is a typical property of most digital systems that they fail abruptly. The aim of this research is to modify a digital system (MPEG-2) so it behaves more like an analogue system, where performance degrades gracefully with increasing channel noise. TRANSCODER APPROACH Standard techniques for improving the MPEG-2 noise performance concentrate on forward error correction (FEC). These techniques add redundancy thereby lowering the coding efficiency. FEC also causes abrupt failure as the error rate increases. Instead we consider the lossless `black box' approach of figure 2 where MPEG-2 data is transcoded into a more resilient structure, transmitted over a noisy channel, and finally recoded back into a compliant MPEG-2 format, to be read by an MPEG-2 decoder. MPEG 2 ENCODER BLACK BOX TRANSCODER INVERSE BLACK BOX MPEG 2 DECODER LOSSY CHANNEL Fig. 2. The MPEG-2 Transcoder The `black box' system is designed to provide substantial resilience to errors. It is lossless, so, in the absence of errors, the output is exactly the same as the input. The system does not significantly alter the bit-rate, which means that the transcoded data can still be sent down the original channel. ERROR PROPAGATION Once an error occurs in MPEG-2, it often causes the decoder to lose bitstream synchronisation. Once synchronisation is lost, all the remaining data is misinterpreted. This point can be shown in the following example: Suppose we have four possible events, A,B,C, and D. Event A is the most probable and C and D are less probable. To achieve compression, we allocate a short code to A and longer codes to B and C. Event A B C D Code 0 11 100 101 Table 1. A Variable Length Code If we now consider a sequence of events, [A,B,A,B,A,A]. This is coded as: [ 0 1 1 0 1 1 0 0]. If we now consider one single error in the second bit, the received signal is : [ 0 0 1 0 1 1 0 0 ]. The received signal is now decoded as [A,A,D,C]. Not only has this one error corrupted everything after it, but the received code contains a different number of symbols (events) from the transmitted code. This means that the decoder has lost symbol-synchronisation, and all the following data is liable to be useless. One solution to this problem is the introduction of re-synchronising codewords. In MPEG-2, a 24-bit code is inserted at the beginning of each slice. Providing the synchronising code is not corrupted, one error will not cause more than once slice to be corrupted. Hence the lost horizontal stripes in figure 1. We conclude that the greatest benefit occurs when bitstream synchronisation in MPEG-2 is obtained at the block layer. However, using traditional techniques, resynchronising as often as this would require such a large coding overhead that this would not be a sensible solution. ERROR RESILIENT SYNCHRONISATION MPEG-2 divides a picture up into 8x8 blocks. These blocks undergo a discrete cosine transformation (DCT). This transform data is then compressed using a variable length code like the example in table 1. The outcome of the compression process is that we have image blocks which are coded using a variable number of bits. To achieve block synchronisation, we must ensure that each variable length block starts at a known position in the transmitted bitstream. We achieve this using a technique known as the error-resilient entropy code (EREC) (2,3). Figure 3a shows N variable length blocks of data. Each block can be considered to be the data required to code an image block, and its height corresponds to the number of bits in that block. The idea behind the EREC is to fit these variable length blocks into the fixed slot structure of figure 3b. AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA A A A A A A A A A A A A A A A A A A A A A A AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAA AAA AAA AAA AAA AAA AAA AAA AAA AAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA Fig. 3a. Variable Length Blocks. The horizontal line in figure 3a corresponds to the average block length. This sets the slot height in figure 3b. Therefore the number of slots is N and the total number of spaces in the slot structure is equal to the total number of bits in all of the variable length blocks. Fig. 3b. EREC Slot Structure. Initially as much as possible of each block of data is placed in the corresponding slot. Any data which overfills its slot is retained as in figure 3c. In the second stage of the EREC encoding process, the remaining data is shifted to the right. Here it attempts to fill any spaces left by smaller than average blocks (figure 3d). AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA Fig. 3c EREC Stage 1. This process then continues: the remaining block data is shifted again, and attempts to fill spaces left in the slot structure. AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AAAA AAAA AAAA AA AA AA AA AAAA AAAA AAAA AAAA Fig. 3d. EREC Stage 2 AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AA AA AA AAAA AAAA AAAA AA AA AA AA AAAA AAAA AAAA AAAA