Методы обеспечения устойчивости к ошибкам

11.5.7 Redundant motion vectors

Many of the error resilience tools present in H.264/AVC have been omitted in subsequent standards. Carreira et al. [19] exposed the increased susceptibility of H.265/HEVC coded streams to network errors and proposed a two-stage approach which limited temporal error propagation following frame loss. Firstly, at the encoder, reference pictures are selected dynamically, using Lagrangian optimization to ensure that the number of prediction units that depend on a single reference is reduced. Secondly, they employ motion vector prioritization, based on spatial dependencies, to select an optimal subset of redundant motion vectors to transmit as side information. These results indicate significant reductions in temporal error propagation, achieving total quality gains up to 5 dB for 10% of packet loss ratio for only a small increase in motion overhead.

1 Introduction

This chapter addresses the cross—layer resilience challenge for designing rugged embedded systems: given a set of resilience techniques at various abstraction layers (circuit, logic, architecture, software, algorithm), how does one protect a given design from radiation-induced soft errors using (perhaps) a combination of these techniques, across multiple abstraction layers, such that overall soft error resilience targets are met at minimal costs (energy, power, execution time, area)? Specific soft error resilience targets include: silent data corruption (SDC), where an error causes the system to output an incorrect result without error indication; and detected but uncorrected error (DUE), where an error is detected (e.g., by a resilience technique or a system crash or hang) but is not recovered automatically without user intervention.

The need for cross—layer resilience, where multiple error resilience techniques from different layers of the system stack cooperate to achieve cost-effective error resilience, is articulated in several publications (e.g., Refs. [1–7]).

CLEAR (cross-layer exploration for architecting resilience) is a first of its kind framework, which addresses the cross-layer resilience challenge. In this chapter, the focus is on the use of CLEAR for harsh-environment-capable embedded systems that operate in the presence of radiation-induced soft errors in terrestrial settings; however, other error sources (voltage noise and circuit-aging) may be incorporated into CLEAR as well.

Although the soft error rate of an SRAM cell or a flip-flop stays roughly constant or even decreases over technology generations, the system-level soft error rate increases with increased integration [12–14]. Moreover, soft error rates can increase when lower supply voltages are used to improve energy efficiency [15,16]. The chapter focuses on flip—flop soft errors because design techniques to protect them are generally expensive. Coding techniques are routinely used for protecting on-chip memories. Combinational logic circuits are significantly less susceptible to soft errors and do not pose a concern [14,17]. Both single-event upsets (SEUs) and single-event multiple upsets (SEMUs) [16,18] are considered. While CLEAR can address soft errors in various digital components of a complex System-on-a-Chip (including uncore components [19] and hardware accelerators), a detailed analysis of soft errors in all these components is beyond the scope of this chapter. Hence, this chapter will focus on soft errors in processor cores.

10.1 Overview

Chapter 9 discussed three categories of error-resilience techniques: forward, postprocessing (or concealment), and interactive techniques. Almost all forward techniques increase the bit rate because they work by adding redundancy to the data, e.g., FEC. Some of them may also require modifications to the encoder, e.g., layered coding, and others may not be suitable for some applications, e.g., multiple description coding assumes several parallel channels between transmitter and receiver. Most interactive techniques depend on a feedback channel between the encoder and decoder. Such a channel may not be available in some applications, e.g., multipoint broadcasting. Most interactive techniques will also introduce some delay and may, therefore, be unsuitable for real-time applications like mobile video communication. On the other hand, concealment techniques do not increase the bit rate, do not require any modifications to the encoder, do not introduce any delay, and can be applied in almost any application. This makes them a very attractive choice for mobile video communication, where bit rate and delay are very critical issues.

A very successful class of error concealment is temporal error concealment. Conventional temporal concealment techniques estimate one concealment displacement for the whole damaged block and then use translational displacement compensation to conceal the block from a reference frame. The main problem with such techniques is that incorrect estimation of the concealment displacement can lead to poor concealment of the whole or most of the block.

This chapter describes the design of two novel temporal concealment techniques. In the first technique, motion field interpolation (MFI) is used to estimate one concealment displacement per pel of the damaged block. Each pel is then concealed individually. In this case, incorrect estimation of a concealment displacement will affect only the corresponding pel. On a block level, this may affect few pels rather than the entire block. In the second technique, multihypothesis motion compensation (MHMC) is used to combine the first technique with a boundary matching (BM) temporal concealment technique to obtain a more robust performance.

The chapter also investigates the performance of different temporal error concealment techniques when incorporated within a multiple-reference video codec. In particular, the chapter finds a combination of techniques that best recovers the spatial-temporal components of a damaged multiple-reference motion vector. In addition, the chapter describes the design of a novel multihypothesis temporal concealment technique that can be used with multiple-reference systems.

The rest of the chapter is organized as follows. Section 10.2 describes the MFI temporal concealment technique, whereas Section 10.3 presents the combined BM-MFI technique. Section 10.4 presents some simulation results. Section 10.5 investigates the performance of temporal error concealment within multiple-reference video codecs. It also describes the multihypothesis multiple-reference temporal concealment technique. The chapter concludes with a discussion in Section 10.6.

Preliminary results of this chapter have appeared in Refs. 204, 205, 206, 207, and 208.

5 Error-Resilient Video Transcoding

With the emergence of various advanced wireless networks such as 3G cellular systems, wireless video has become increasingly popular and is attracting great interest [38–40]. Transmitting coded video stream over wireless channels is even more challenging than over wired channels since wireless channels usually have lower bandwidth and higher bit error rates than wired channels.

In [7], the authors presented a novel fully comprehensive mobile video communication system that exploits the useful rate management features of the video transcoders and combines them with error resilience for transmissions of coded video streams over general packet radio service (GPRS) mobile access networks. The error-resilient video transcoding operation takes place in a video proxy, which not only performs bit rate adaptation but also increases error resilience. In the proposed system, two resilience schemes are used, which are adaptive intrarefresh (AIR) and feedback control signaling (FCS) methods. These two schemes can work independently or combined. The system adjusts the output bit rate from the proxy by monitoring the current channel conditions such as bit error rate, delay, and so forth. Meanwhile, the amount of resilience added to the video data can also be controlled by monitoring the proxy output rate and the change in the error conditions of the network connection. AIR is mainly used to stop the error propagation across different frames. Experiments showed that the combined AIR-FCS method gave superior transcoding performances over error-prone GPRS channels relative to nonresilient video streams.

H.263 extensions (H.263+ and H.263++)

H.263 Version 2, also known as H.263+, was standardized in 1998 and extends H.263 with many new modes and features that further improve compression efficiency. H.263++ (Version 3) provided still further enhancements in 2000. The 12 new modes in H.263+ improve coding gain, improve error resilience, enable scalable bitstreams, introduce flexibility in picture size and clock frequency, and provide supplemental display capabilities. Many of these new modes were pulled through into the later H.264/AVC standard.

9.6.3.2 The Error-Resilience Entropy Code (EREC)

An interesting alternative to inserting resynchronisation codewords is the error resilience entropy code (EREC) [177, 185]. The EREC takes variable-length blocks of data and rearranges them into fixed-length slots. For example, assume that there are N variable-length blocks with lengths b_i, i = 1 … N. The encoder first chooses a total data size T ≥ Σ b_i, which is sufficient to encode all the data. This total data size is split into N slots of fixed lengths s_i, i = 1 … N. An N-stage algorithm is then used to place the data from the variable-length blocks into the fixed-length slots. At each stage n, a block i with data left unplaced searches slot j = i + ϕ_n (mod N) for space to place some or all of the remaining data. Here, ϕ_n is an offset sequence that is usually pseudo-random.

In the presence of errors, the resilience provided by the EREC algorithm is due to two factors. First, each block starts at a known position in the bitstream (i.e., the start of the corresponding slot). Thus, in the case of loss of synchronization, the decoder simply jumps to the start of the next slot without the need for resynchronization codewords. Second, subjectively less important data (e.g., high-frequency DCT coefficients) are usually placed in later stages of the algorithm. With the EREC algorithm, most error propagation effects (due, for example, to missing or falsely detecting end-of-block codewords) hit data placed at later stages of the algorithm rather than the more important data at the start of the slots.

C. Coding Efficiency

Besides the provision for new content-based functionalities and error resilience and robustness, the coding of video with very high coding efficiency over a range of bit rates continues to be supported for the MPEG-4 standard. As indicated above, the MPEG-4 Video VM allows the single object-layer (single VOP) coding approach as a special case. In this coding mode, the single VOP input image sequence format may be rectangular as depicted in Fig. 7 (thus not segmented into several VOP’s), and the MPEG-4 Video VM coding algorithm can be made almost compatible to the ITU-H.263 or ISO-MPEG-1 standards. Most of the coding techniques used by the MPEG-2 standard at Main Profile are also supported. A number of motion compensation and texture coding techniques are being investigated in the Core Experiment process to further improve coding efficiency for a range of bit rates, including bit rates below 64 kb/s.

6.4.5 Objective Adjustment of the Efficiency-Complexity Tradeoff

In sequential algorithms, the decoding performance in terms of error resilience efficiency is improved when more paths are explored in the decoding tree; however, this improvement goes with increased decoding complexity. To adjust a tradeoff between decoding complexity and efficiency, this section introduces some test to decide whether a path is worth being expanded by the MA or SA.

Consider a path of v – 1 bits corresponding to b_1:v–1. In a bit-by-bit exploration, this path may be extended with b_v = 0 and b_v = 1. Assume that both extended paths correspond to prefixes of sequences in Ω_n. A hard decoder discards systematically one of them and, thus, explores only one path in the decoding tree. The idea is to derive a test allowing to decide if both of the two choices deserve being considered or if a hard decision on the current bit is sufficient.

3.1.7 Error Resilience

To ensure robust operation over error-prone channels, MPEG-4 Part 2 offers a suite of error-resilience tools that can be divided into three groups: resynchronization, data partitioning, and data recovery [8, 22]. Resynchronization is enabled by the MPEG-4 Part 2 syntax, which supports a video packet structure that contains resynchronization markers and information such as macroblock number and quantizer in the header. All of these are necessary to restart the decoding operation in case an error is encountered. Data partitioning allows the separation between the motion and texture data, along with additional resynchronization markers in the bitstream to improve the ability to localize the errors. This technique provides enhanced concealment capabilities. For example, if texture information is lost, motion information can be used to conceal the errors. Data recovery is supported in MPEG-4 Part 2 by reversible variable-length codes for DCT coefficients and a technique known as new prediction (NEWPRED). Reversible variable length codes for DCT coefficients can be decoded in forward and backward directions. Thus, if part of a bitstream cannot be decoded in the forward direction due to errors, some of the coefficient values can be recovered by decoding the coefficient part of the bitstream in the backward direction. NEWPRED, which is a method intended for real-time encoding applications, makes use of an upstream channel from decoder to encoder, where the encoder dynamically replaces the reference pictures according to the error conditions and feedback received from the decoder [6].