Feedback priority control for MPEG video transmitted through ATM networks

Abstract

Although ATM networks are suitable for transporting variable bit rate (VBR) traffic, bursty moving picture exports group (MPEG) traffic is tending to make the networks congested and results in cell loss. It is suggested that cells should be assigned different priorities according to the significance of the contained information such that the networks can try to protect the high-priority cells and obtain high video quality. Considering the payload behind the slice header, losing any cell would disable the decoding for the following part of the same slice. As a result, if any cell has to be discarded, the farther the position is, the lesser the following data are affected. In this article, a simple mechanism called ‘feedback control priority partition with RC-shaper (FBCPP&RC)’ is proposed to prevent losing cells at the former position by adaptively assigning high priority to these cells. The decision policy is simple and quick. It also employs an RC-shaper to smooth the burstiness of the cells so as to reduce cell-loss probability. The mechanism is expected to facilitate transmitting MPEG video as a cheaper available bit rate (ABR) traffic rather than VBR traffic. According to the simulation results, it performs well in improving the quality of single-frame as well as continuous pictures. It can also be applied to the new standard MPEG-II.

Introduction

Moving picture exports group (MPEG) is a standard for compressing digital full-motion video to a ratio of approximately 1/100 [1]. An MPEG encoder compresses the video data into a coded bit-stream representing a sequence of frames (pictures). There are three types of encoded frames, I (intra-coded), P (predicted), and B (bi-directional). Because MPEG traffic is bursty, time-sensitive and has a high bit rate, it is suggested that it is transmitted over ATM networks.

The errors that occur in ATM networks can be classified into bit errors and cell loss. When an optical fiber is employed, the bit error probability is negligible. As a result, the major error in ATM networks is cell loss within switches and multiplexers as a result of congestion, or within enforcers as a result of violating the declared traffic parameters. Being bursty, the MPEG traffic tends to violate the declared parameters, making the networks congested and therefore resulting in cell loss. In order to protect the quality of the video traffic, some schemes were proposed to resolve the burstiness of the traffic.

One of these is smoothing the bursty traffic to a constant bit rate (CBR) or near-CBR traffic by employing a buffer or shaper [2], [3], [4]. Such schemes require large buffers to store the waiting cells. As a result, Ouyang [5] proposed a ‘multiple leaky buckets (MLB)’ scheme, where the leak rate and buffer capacity of all LBs are shared in the same virtual path so as to reduce the delay and improve the resource utilization of networks. However, the delay variation introduced in the buffer would still degrade the video quality. On the contrary, if the original traffic was not smoothed, not only would the video quality be degraded owing to cell loss, but the network resource would also be wasted owing to re-transmission of the data of other connections. [6], [7], [8]. Another way is to adjust the traffic rate by changing the MPEG compressing parameters N (which is the minimal number of frames between two I-frames) and M (which is the minimal number of frames between a P-frame and an I-frame or another P-frame) [9]. Because I- and P-frames contain a much higher data rate compared with that of B-frames, enlarging the two parameters would reduce the frequency of generating such frames. The traffic rate is reduced as a result.

Apart from the problem of burstiness, the dependence between frames arising from the compressing technique is also a problem. Because both I-frames and P-frames are referred by B-frames (an I-frame is also referred by a P-frame), once an error occurs in such frames, the error would be propagated to the following related frames. As a result, the sequence of the significance of the frame types is I, P and B. Considering the encoded bit-stream within a frame, the headers (such as sequence header, GOP header, picture header and slice header) necessary for decoding are most important. If any cell of the headers is lost, the resulting error propagation would be much more severe compared with that of the data part. As a result, it is suggested that cells be assigned to different priorities depending on the frame types as well as the significance of the contents (i.e. whether header or data). With such pre-processing, the network can try to protect the high-priority cells and maintain the quality of the video. Usually, all the headers are set as high priority because they are necessary for decoding. As for the data part, Ismail [10] suggested assigning high priority to the cells carrying low-frequency direct cosine transform (DCT) coefficients, while assigning low priority to the cells carrying high-frequency coefficients. However, some extra modules, such as variable length coding (VLC) decoder, encoder and quantizer must be attached with the MPEG encoder to realize such a scheme.

Because MPEG belongs to VLC, losing any cell of the data part would disable the decoding for the remaining part of the same slice. Compared with the cells corresponding to the latter position, the cells located at the former position of a slice are more important and should be assigned a high priority. Based on such a principle, we propose a mechanism, namely “feedback control priority partition with RC-shaper (FBCPP&RC),” to determine the priority of a cell depending on the position of the cell as well as the network condition. This mechanism is composed of a “feedback control priority partition (FBCPP)” scheme and an RC-shaper, where the FBCPP scheme is designed for determining the priorities of cells; while the RC-shaper is used to smooth the burstiness of the coded bit-stream.

As usual, the performance of the proposed mechanism is evaluated based on the PSNR for compressed pictures rather than on the cell-loss rate for non-compressed data. The PSNR is defined as

$$\text{PSNR}\text{=10}\text{log}\text{(255}{}^2\text{/}\text{MSE}\text{)}$$

with mean square error, $\text{MSE}\text{={}\text{∑}\text{i=1}\text{352}\text{∑}\text{j=1}\text{240}\text{(S}{}_{\mathrm{ij}}\text{−P}{}_{\mathrm{ij}}\text{)}{}^2\text{}/(352×240)}$ where S_ij is the pixel value of the original frame before encoding, while P_ij is that of the frame received at the destination (assuming that one frame includes 352×240 pixels) [1]. Note that because the MPEG compressing technique will intrinsically result in distortion, the MSE is never equal to zero even when no cell is lost. As a result, the value of PSNR will be finite.

In the next section, the proposed mechanism is presented. The performance evaluated by simulation is illustrated in Section 3. Finally, some discussions and conclusions are drawn in the last section.