Stochastic channel-adaptive rate control for wireless video transmission

https://doi.org/10.1016/j.patrec.2004.01.003Get rights and content

Abstract

In this paper, an empirically optimized channel-matched quantizer, and a joint stochastic-control based rate controller and channel estimator for H.261 based video transmission over a noisy channel is proposed. The rate controller adaptively learns to choose the correct channel matched quantizer using a stochastic learning algorithm. The stochastic automaton based learning algorithm aids in estimating the channel bit error rate based on a one bit feedback from the decoder. The algorithm is observed to converge to the optimal choice of the quantizer very quickly for various channel bit error probabilities and for different video sequences. When compared to traditional channel estimation schemes the proposed technique has several advantages. First, the proposed method results in a significant reduction in the delay and bandwidth requirement for channel estimation when compared to pilot symbol aided channel estimation schemes. Next, the stochastic learning algorithm used to estimate the channel bit error rate has simple computations. This makes it attractive for low power applications such as wireless video communications. This is in contrast to traditional blind channel estimation schemes that are computationally expensive, in general.

Introduction

Multimedia applications such as the wireless video transmission have lead to the study of issues in error-resilient low bit-rate video transmission over noisy channels. Wireless links not only suffer from limited bandwidth problems but are also highly vulnerable to channel errors. Video compression standards like the H.261 (Bhaskaran and Konstantinides, 1995) alleviate the bandwidth problem to a certain extent. The H.261 standard also known as the p×64 standard was developed for video coding and decoding at the rate of p×64 kbits/s, where p is an integer from 1 to 30.

Most of the state-of-the-art video codecs treat source and channel coding separately. Bandwidth reduction is achieved by the source coder by removing the redundancy in the source statistics. Error protection against channel error is take care of by the channel coder through the addition of redundancy in the transmitted data. However, this separation is justifiable only in the limit of an arbitrary encoding/decoding complexity. But, we know that in practice complexity and delay are the main constraints for communication systems. Therefore, the separation of source and channel coding is no longer optimal. This implies that source and channel coding should depend on each other leading to joint source–channel coding (JSCC) (Kurtenbach and Wintz, 1969).

JSCC has been receiving significant attention lately as a viable solution for achieving reliable communication of signals across noisy channels. The rationale behind using such techniques is the observation that Shannon’s source–channel separation theorem (Shannon, 1949) does not usually hold under delay and complexity constraints or for all channels (Vembu et al., 1995). JSCC tries to design the source coder and channel coder in some joint way, which can provide better error protection and bandwidth utilization. JSCC schemes can be broadly classified into three different categories, joint source–channel encoding (JSCE) (Dunham and Gray, 1981; Farvardin, 1990; Phamdo et al., 1997), joint source–channel decoding (JSCD) (Sayood and Borkenhagen, 1991; Sayood et al., 1994; Phamdo and Farvardin, 1994; Park and Miller, 1997, Park and Miller, 2000; Demir and Sayood, 1998; Wen and Villasenor, 1999; Bauer and Hagenauer, 2000a, Bauer and Hagenauer, 2000b; Hedayat and Nosratinia, 2002; Kliewer and Thobaben, 2002; Guivarch et al., 2000; Subbalakshmi and Vaisey, 1998, Subbalakshmi and Vaisey, 1999a, Subbalakshmi and Vaisey, 1999b, Subbalakshmi and Vaisey, 2001, Subbalakshmi and Vaisey, in press; Murad and Fuja, 1998a, Murad and Fuja, 1998b; Lakovic et al., 1999; Lakovic and Villasenor, 2002; Alajaji et al., 1996; Burlina and Alajaji, 1998; Kopansky and Bystrom, 1999; Bystrom et al., 2001; Subbalakshmi and Chen, 2002; Chen and Subbalakshmi, 2003) and rate allocation strategies (Hochwald and Zeger, 1997; Bystrom and Modestino, 1998; Cheung and Zakhor, 2000). As the names suggest, these deal with the joint design of encoders, decoders and the rate allocation between the channel and source codes respectively. One early work in this class is by Dunham and Gray (1981), where they demonstrate the existence of a joint source–channel system for special source and channel pair, by showing that a communication system using trellis encoding of a stationary, ergodic source over a discrete memoryless noisy channel can perform arbitrarily close to the source distortion-rate function evaluated at the channel capacity. Other works include an index assignment algorithm proposed for the optimal vector quantizer on a noisy channel (Farvardin, 1990) and the design of quantizers for memoryless and Gauss–Markov sources over binary Markov channels (Phamdo et al., 1997).

Work on rate allocation between the channel and source codes includes the optimal allocation algorithm between a vector quantizer and a channel coder for transmission over a binary symmetric channel (BSC) (Hochwald and Zeger, 1997), the optimal source–channel rate allocation to transmit H.263 coded video with trellis-coded modulation over a slow fading Rician channel (Bystrom and Modestino, 1998) and an algorithm to distribute the available source and channel coding bits among the sub-bands of scalable video transmitted over BSC to minimize the expected distortion (Cheung and Zakhor, 2000).

JSCD schemes can be further classified into constrained JSCDs and integrated JSCDs. Constrained JSCDs are typically source decoders that are built using prior knowledge of channel characteristics while integrated JSCDs combine the source and channel decoder into one unit. One example of constrained JSCD for fixed length encoded sources is the work of Sayood and Borkenhagen (1991), who investigated the use of residual redundancy left in the source after coding it with a differential pulse code modulation (DPCM) source coder in providing error protection over a BSC. This was then extended to include conventional source coder/convolutional coder combinations (Sayood et al., 1994). Other work in this class includes the design of a MAP detector for fixed length encoded binary Markov source over a BSC (Phamdo and Farvardin, 1994) and a MAP decoder for hidden Markov source (Park and Miller, 1997). Channel-matched source rate control or quantization has been shown to be an effective way to add error-resilience to the transmission of compressed images and video over noisy channels (Kurtenbach and Wintz, 1969; Shannon, 1949; Vembu et al., 1995; Dunham and Gray, 1981; Farvardin, 1990; Phamdo et al., 1997; Sayood and Borkenhagen, 1991; Sayood et al., 1994; Phamdo and Farvardin, 1994; Park and Miller, 1997, Park and Miller, 2000; Demir and Sayood, 1998; Wen and Villasenor, 1999; Bauer and Hagenauer, 2000a, Bauer and Hagenauer, 2000b; Hedayat and Nosratinia, 2002; Kliewer and Thobaben, 2002; Guivarch et al., 2000; Subbalakshmi and Vaisey, 1998, Subbalakshmi and Vaisey, 1999a, Subbalakshmi and Vaisey, 1999b, Subbalakshmi and Vaisey, 2001, Subbalakshmi and Vaisey, in press; Murad and Fuja, 1998a, Murad and Fuja, 1998b; Lakovic et al., 1999; Lakovic and Villasenor, 2002; Alajaji and Fuja, 1994; Alajaji et al., 1996; Burlina and Alajaji, 1998; Kopansky and Bystrom, 1999; Bystrom et al., 2001; Fano, 1963; Subbalakshmi and Chen, 2002; Chen and Subbalakshmi, 2003; Hochwald and Zeger, 1997; Bystrom and Modestino, 1998; Cheung and Zakhor, 2000; Chandramouli et al., 1998a).

The H.261 standard recommends the use of Huffman encoding to achieve an additional gain in the compression ratio. But, it is known that variable length codes are highly susceptible to channel errors. The critical bits need to be protected from channel errors in order to prevent the complete loss of a transmitted video sequence. If, during transmission some bits are flipped, added or dropped, the synchronization of the decoder to the received bit stream could be lost. This leads to error propagation and the loss of the source symbols. The loss of a few blocks of symbols causes displacements in the received image. Error correcting codes can be used to protect the critical bits from channel errors. Examples of the critical bits are the EOB (end of block) markers and the most significant bit of a source symbol. An error in the most significant bit could cause higher degradation than a corrupted least significant bit. The loss of EOB due to errors leads to catastrophic error propagation as shown in Fig. 1. Therefore, the high priority bits need to be protected using channel coding or other methods. But the redundancy due to channel coding reduces the effect of the compression efficiency. Therefore, an optimal trade-off between the rate of the source coder and the channel coder is essential.

Channel-matched source quantization and adaptive source rate control based on the channel characteristics are effective ways of reducing the effects of channel noise on the received video signal. However, the performance depends on how fast and reliably the channel parameters (such as the bit error probability pe) can be estimated. In many applications, pe is computed using pilot symbol aided techniques. This causes large delays which may not be acceptable for real-time applications. It has also been observed that up to a 14% loss in capacity can be incurred due to pilot symboling (Cavers, 1991). Therefore, it is desirable to reliably estimate the channel statistics and also achieve rate control through adaptive quantization on the fly with minimal overhead.

In this paper, a channel-matched quantizer, and a fast and reliable simultaneous rate control and channel estimation algorithm based on the stochastic learning automaton (Narendra and Thathachar, 1989) for a H.261 based video codec over channels that cause random bit errors is proposed. A stochastic learning automaton at the encoder estimates and tracks the channel bit error probability. For simplicity, we only consider pe=10−1, 10−2 and 10−3 only in this work as they are typical of the wireless channels. However, we note that this method can be extended to finite state channel models with more number of states by expanding the action set of the learning automaton. The learning is based on a one bit decision feedback from the decoder summarizing the peak signal to noise ratio (PSNR) of the received video frames for a particular choice of the source quantizer. The optimal quantizer (and hence the bit rate due to the one-to-one mapping) for that channel bit error probability is learnt and selected by the learning automaton using a linear reward inaction (LRI) learning scheme. We realized that the LRI scheme has absorbing barriers (Narendra and Thathachar, 1989). One way to overcome this would be to trigger the learning process if the received video quality is consistently below a certain threshold. There is no additional overhead of pilot symbols in the proposed approach. We also note that the proposed technique is particularly suited for low power wireless video applications where computationally simplicity in encoding and decoding is emphasized. The system designer has the flexibility to control the convergence rate of the learning algorithm depending on the reliability and delay constraints. To our knowledge, this is the first attempt in using stochastic learning automaton for rate control in low-bit rate video transmission. In order to prevent sync losses at the decoder, the fast error resilient entropy code in (Chandramouli et al., 1998b) is also used. We organize the paper as follows. In Section 2 the channel-matched quantization technique is discussed. The variable structure stochastic learning automaton is introduced in Section 2.3 followed by the rate control and channel estimation algorithm. Performance of the proposed algorithm is studied in Section 3 and concluding remarks are given in Section 4.

Section snippets

Quantizer design

The proposed H.261 based video codec is shown in Fig. 2. The adaptive quantizer is implemented using a VSLA. This will be discussed subsequently. To prevent synchronization loss due to error propagation in the variable length coded transmitted data a error-resilience code called FEREC (Chandramouli et al., 1998c) is used. We now discuss a channel matched source quantization scheme. This is similar to the one proposed for the transmission of JPEG compressed images (Chandramouli et al., 1998c).

Performance analysis

In this section we discuss the performance of the adaptive quantizer and on-line channel estimator. The H.261 video codec used for experiments was implemented in C by the authors and their students. Results are presented for the standard grey level, qcif format, Miss America video sequence at 10 frames/s. The uncompressed bit rate was about 2 Mbps and the highest compressed video sequence had a rate of 35 Kbps for a particular choice of the quantization factor given in the following

Conclusion

A joint adaptive rate control and channel estimation algorithm based on stochastic learning is presented. First, channel-matched quantizer design is discussed. The quantizers are optimized empirically based on simulation data. The application of LRI learning to source rate control for video transmission is studied. Some convergence properties of this method are analyzed. The convergence of the algorithm to the optimal channel matched quantizer is fast. The learning delay and optimal

Acknowledgements

This work was partially supported by grants NSF ITR-0082064, NSF CAREER ANI 0133761, and Stevens Center for Wireless Network Security. The authors thank the editors for their comments that helped to improve the presentation of this paper. The authors also thank S. Kumar and Q. Chen for their help.

References (45)

  • F. Alajaji et al.

    A communication channel modeled on contagion

    IEEE Trans. Inform. Theory

    (1994)
  • F. Alajaji et al.

    Detection of binary Markov sources over channels with additive Markov noise

    IEEE Trans. Inform. Theory

    (1996)
  • Bauer, R., Hagenauer, J., 2000a. Symbol by symbol MAP decoding of variable length codes. 3rd ITG Conference Source and...
  • Bauer, R., Hagenauer, J., 2000b. Iterative source/channel-decoding using reversible variable length codes. IEEE Data...
  • V. Bhaskaran et al.

    Image and Video Compression Standard

    (1995)
  • J. Borko et al.

    Motion Compensation Algorithms for Video Compression Standard

    (1997)
  • P. Burlina et al.

    An error resilient scheme for image transmission over noisy channels with memory

    IEEE Trans. Image Process.

    (1998)
  • Bystrom, M., Modestino, J.W., 1998. Combined source–channel coding for transmission of H.263 coded video with...
  • M. Bystrom et al.

    Soft source decoding with applications

    IEEE Trans. Circuits Syst. Video Technol.

    (2001)
  • J. Cavers

    An analysis of pilot symbol assisted modulation for rayleigh fading channels

    IEEE Trans. Vehicular Technol.

    (1991)
  • R. Chandramouli et al.

    Joint optimization of quantization and on-line channel estimation for low bit-rate video transmission

    Proc. IEEE Int. Conf. Image Process.

    (1998)
  • Chandramouli, R., Ranganathan, N., Ramadoss, S., 1998b. Error resilient coding for jpeg image transmission over...
  • R. Chandramouli et al.

    Adaptive quantization and fast error resilient entropy coding for image transmission

    IEEE Trans. Circuits Syst. Video Technol.

    (1998)
  • Q. Chen et al.

    Trellis decoding for MPEG-4 streams over wireless channels

    SPIE Electr. Imaging: Image Video Commun. Process.

    (2003)
  • G. Cheung et al.

    Bit allocation for joint source/channel coding of scalable video

    IEEE Trans. Image Process.

    (2000)
  • Demir, N., Sayood, K., 1998. Joint source/channel coding for variable length codes. IEEE Data Compression Conference,...
  • J.G. Dunham et al.

    Joint source and noisy channel trellis encoding

    IEEE Trans. Inform. Theory

    (1981)
  • R. Fano

    A heuristic discussion of probabilistic decoding

    IEEE Trans. Inform. Theory

    (1963)
  • N. Farvardin

    A study of vector quantization for noisy channels

    IEEE Trans. Inform. Theory

    (1990)
  • Guivarch, L., Carlach, J.-C., Siohan, P. 2000. Joint source–channel soft decoding of Huffman codes with turbo-codes....
  • Hedayat, A., Nosratinia, A., 2002. On joint iterative decoding of variable-length codes and channel codes. IEEE...
  • B. Hochwald et al.

    Tradeoff between source and channel coding

    IEEE Trans. Inform. Theory

    (1997)
  • Cited by (7)

    View all citing articles on Scopus
    View full text