KR100882759B1 - Method and apparatus for joint source-channel map decoding - Google Patents

Abstract

Receiving a bitstream comprising one or more bits, determining whether the bitstream has one or more corrupted bits, determining one or more hypotheses representing an error pattern, and assigning probabilities to each of the hypotheses And the probability is determined based on one or more reference data.

Error correction, joint source-channel, MAP decoding.

Description

METHOD AND APPARATUS FOR JOINT SOURCE-CHANNEL MAP DECODING}

35 U.S.C.§ Of 119  Priority claim

This patent application claims priority to US Provisional Application No. 60 / 569,400, filed May 6, 2004, entitled "Method and Apparatus for Joint Source-Channel MAP VLC," to the assignee of the present invention. Is hereby expressly incorporated by reference herein.

Background of the Invention

Field

The present invention relates to digital communications, and more particularly to channel and source decoding.

Background

The demand for higher data rates and higher quality of services in mobile communication systems is growing rapidly. However, factors such as limited transmit power, limited bandwidth and multi-path fading continue to limit the data rates handled by the actual system. In multimedia communications, especially in error-prone environments, even errors in one decoded value can lead to decoding artifacts that propagate spatially and temporally, so that the error resilience of the transmitted media is of a certain quality. It is crucial in providing a service. Various encoding means have been used to minimize the error while maintaining the necessary data rate, but all these techniques suffer from the problem of arriving at the decoder side.

Through the use of a source encoder, the data is compressed, extending the minimum number of bits to convey maximum information, followed by a channel encoder that tends to maximize the capacity of the transmission channel for a given probability of error when receiving such bits. Connected.

Channel coding, eg Reed-Solomon coding, is used to improve the robustness of the source coded data. The joint source-channel coding methodology provides non-uniform error protection for coded source data with non-uniform importance or makes the rate suitable for coded video data in the available network bandwidth through segmentation and dropping of packets. This is because conventional transport protocols do not deliver corrupted data to the source decoder.

Source coding techniques such as reversible variable length coding (eg, in MPEG-4) have been used for error recovery by decoding packets in reverse order when the corrupted packet is actually received. There is a compromise in coding efficiency with source coding techniques that translate to the quality of decoded video at a given bit rate.

Entropy coding enables a highly efficient lossless representation of the symbols generated by random information sources. As such, it is an essential component in both lossless and lossy data compression schemes. Despite its tremendous benefit in compression efficiency, entropy coding also complicates the decoding process. A conventional feature of all other approaches to entropy coding is that one or consecutive source symbols (dictionary coding) are concatenated and represented in a binary pattern, ie a sequence of 1's and 0's known as codewords. The length increases with decreasing approximation. Thus, the symbols are assigned to smaller, more economical representations, and on average allow significant savings on fixed length representations based on direct symbol alphabetic size.

Uncertainty in the representation of how many bits to consume for the next symbol in the bitstream, ie the entropy coded representation of the output of the information source, is an obvious complexity for the decoder. However, more importantly, in the case of errors in the bitstream, the use of codewords of various sizes in combination with flipped bits (due to errors) will frequently result in emulation of incorrect codeword lengths. As a result the parsing / decoding process will lose synchronization with the bitstream, i.e. correct the correct identification of the codeword boundary and thus the correct interpretation of the bitstream will begin to fail.

Assume that a decoder implementing a basic level of error detection measurement faces problems in decoding the bitstream and loses synchronization. Eventually, due to syntax violation, i.e. an invalid codeword, or semantic failure, e.g. an invalid parameter value or an unexpected bitstream object, the decoder will recognize the problem and resolve the bitstream and itself. You will take the necessary steps to resynchronize. This will typically result in data loss to a degree that goes beyond the corruption that caused the data loss in the first location. One reason for this is the fact that the points in the bitstream that enable resynchronization are not frequently used due to the aspect ratio that provides them, with slice boundaries being the most common example. More importantly, across these points, all forms of encoding / decoding dependencies are avoided, for example resulting from the requirement that the predictive coding chain is terminated and restarted after default initialization. It is a sacrifice of compression efficiency.

Another reason to cause data loss beyond initial corruption is due to incorrect codeword emulation. The identification of the initial bit error location is not a trivial task and is typically not possible without the special design at the bottom of the application layer, namely the channel decoder supporting it. Thus, in detecting bitstream corruption, the source decoder must stop decoding and move forward in the bitstream to find the next resynchronization point, and in this process must overlook a significant amount of potentially healthy data. Emulation of different codewords that are the same length as the original, i.e., reliable codeword, may seem a small problem in relation to the sequencing of the events described above, but in practice it is not. There are a number of ways in which this kind of error may lead to failure in the correct bitstream interpretation of the decoder. For example, in the most recent codec there is an object (compression related parameter) in the bitstream whose value affects the syntax of the following portion of the bitstream. Thus, incorrect values of these objects will lead to incorrect bitstream interpretation.

Even after the introduction of the video compression standard H.264, there are still other reasons for making the emulation of codewords of inaccurate but equal length as poor as synchronization loss. The so-called context adaptation (or context dependency) of entropy coding rules is a situation where the loss of bitstream synchronization is not due to the emulation of codewords of different sizes, i.e. the size of the codewords remains the same, but inaccurate values (incorrect code Decoded from a word may cause an unreliable context change, resulting in a situation that replaces entropy decoding rules as a result of a set of alternative and incorrect rules. In all the cases described, even one incorrectly decoded value will lead to decoding artifacts.

An improved method of correcting bit errors is needed to avoid data loss combined with the synchronization loss of the source decoder and the resulting resynchronization in the entropy coded bitstream. Residual bit errors, i.e. bit errors not detected or detected but not corrected by the used forward error correction scheme, need to be addressed for their decreasing impact on the decoding of the entropy coded bitstream.

summary

A method and apparatus for digital communication that provides improved error correction capability in a decoder such as in a mobile device is described. The noise data received at the mobile device is demodulated and sent to the decoder of the physical layer, such as the associated decoder. After decoding, both the correctly received symbol and the incorrectly received symbol (including one or more corrupted bits) may cause the software application or application specific hardware to be, for example, within the Maximum a Posteriori Probability (MAP) framework. We perform an optimization problem formulation in and solve it to determine the approximation to the hypothesis for entropy coded symbols whose representation is corrupted by the wrong bits.

Correlation information of correctly coded symbols is maintained in one or more memory modules associated with the MAP application. Channel data from the associated channel decoder is monitored and used to account for the various channel conditions and generate a probability distribution function associated therewith. One or more such PDFs are currently available for MAP applications that conform to (actual or measured) channel conditions. Information about bits that are suspected of corrupting incorrectly received symbols is stored in the memory module along with their location information. Information from correctly decoded symbols that may correlate incorrectly received symbols is maintained. The correlation information is used to generate a refined posterior probability distribution. Determining the MAP formulation for an incorrectly received symbol is accomplished by using a properly refined dictionary PDF for that symbol with a conditional PDF of corresponding channel observations conditioned on the hypothesis for the incorrectly received symbol. The result is a MAP function that is maximized to determine the exact value of an incorrectly received symbol.

Brief description of the drawings

1 is an example of a symbol output from a source coder and formatted into a code block for Reed-Solomon erasure coding before turbo coding;

2 is an example of an associated Turbo / Reed-Solomon decoding scheme;

3 is an example of a description of a neighboring 4x4 block that provides correlation data for a corrupted block;

4 is an example of a block diagram of a joint source-channel MAP entropy decoder;

5 is an example of description of consecutive frames;

6 is an example of a flow diagram for a method of joint source-channel MAP entropy decoding.

Detailed description of the invention

In the following description, specific details are given to provide a thorough understanding of the described embodiments. However, it will be understood by those skilled in the art that embodiments may be practiced without these specific details. For example, electronic components may be shown in block diagrams in order not to obscure the embodiments in unnecessary detail. In other instances, such components, other structures and techniques may be shown in detail to further illustrate the embodiment.

It is also noted that an embodiment may be described as a process depicted in a flowchart, flow diagram, structure diagram, or block diagram. Although a flowchart may describe the operations as a sequential process, many of the operations can be performed in parallel or concurrently, and the process can be repeated. In addition, the order of the operations may be rearranged. The process is terminated when its operation is completed. Processes may correspond to methods, functions, procedures, subroutines, subprograms, and the like.

In a digital communication system for compressed data transmission, error recovery methods and apparatus are described as recovering corrupted portions of received data using information received without errors. Data corrupted in an incorrectly received symbol is estimated by statistically evaluating a received symbol with corrupted data and correlated reference data. In one example, the maximum posterior probability (MAP) optimization formulation is performed to manipulate incorrectly received packets at the application layer. To estimate the value of a symbol within this incorrectly received packet, the calculation of the MAP function uses information from the received symbol related to the reference data that correlates to the symbol corrupted by the wrong bit in the incorrectly received packet.

In one embodiment, the MAP framework utilizes statistical analysis of reference data obtained from incoming channel data monitoring. In one example, the method and apparatus may be applied as an extension to the existing joint video team (JVT) H.264 video coding standard, although the basic concept is that error code differs in coded video and / or audio or other form of information. It can be extended to other compressed video, audio, image and general multimedia communication applications that are transmitted beyond a channel that introduces data to data that is attached to different portions of the data or to their correlated portions.

1 is an example of a symbol organized for Reed Solomon erasure and turbo coding output and associated from an H.264 encoder. With respect to this encoding side, the symbols output from the information source, for example the binary codewords output from the H.264 encoder, are blocked into bytes 102. Each byte 102 is considered a symbol in a finite field known as a "Galois Field 256" for the purpose of the outer (N, K) Reed-Solomon (RS) code above the GF 256. . N and K respectively indicate its source data 106 which holds a systematic part in the total RS codeword 104 and the number of symbols, so (NK) is a parity symbol added to each codeword 104 Provide the number of 108. The (N, K) RS code can correct the (N-K) deletion.

The top K column 106 necessarily holds the symbols output from the information source and these symbols can be scanned into the K column in a column-first or row-first manner. Interleaving achieved by row-first scan results in a fairly short burst length. Therefore, row-direction symbol placement is more desirable for use in conjunction with this example. After this initial step of source data placement, by adding NK parity bytes, each L row 104 (of K bytes) is RS coded into N bytes and thus, K + 1, ..., N columns of FIG. 108 is generated. The best K column composed of source data 106 is referred to as an RS information block, and the entire set of N columns is referred to as an RS encoded block or simply as code block 110.

Each column 112 is appended with a trace bit and a CRC (cyclic redundancy check) checksum required for correct operation of the turbo coder. By attaching checksums to each column 112, those columns that fail to satisfy their respective checksums after turbo decoding will be declared deleted. Each code block 110 has one column 112 input to the turbo coder at a time, so that each column is referred to as a turbo information packet. The output bits of the turbo coder are modulated and input into the channel.

Reed-Solomon coding uses an algorithm that relies on the special characteristics of Galois field (GF) operation. Where a set of real numbers can be named as a field and the field has a finite number of elements, this finite field is also called a gallois field. If p is prime and q is any power of p, then there is a code with a symbol from gallois field GF (q). Reed-Solomon codes with code symbols from gallo s field GF (2 ^m ), ie p = 2, and q = 2 ^m are frequently considered in practical digital communication systems and the generator polynomials of this code are Galois field GF It is specified in terms of its route from (2 ^m ).

2 is an example of a turbo / lead-solomon coupled decoding scheme. The error correction scheme adopted uses both the associated code 200, i.e., the inner (channel) code 202 and the outer (channel) code 204. The associated channel decoder 200 consists of a turbo (internal) decoder 202 performing at the physical layer and a Reed-Solomon (external) erasure correction decoder 204 located at the transport / MAC layer. With this associated system 200, noise modulated data 206 is received across the digital communication system 208. Data 206 is received at demodulator 210 with demodulated data 212 passed to turbo decoder 202 to provide error control. The turbo decoder 202 then submits the hard-decision data 214 to an external Reed-Solomon decoder 204 which further reduces the rate of residual errors. Depending on the total number of erases in the code block and the (N-K) number of parity symbols per RS codeword used, the erase correction will succeed or fail. And the data 216 that was correctly received, incorrectly received but corrected, and remained incorrectly received and could not be corrected is passed to the application layer 218.

In the channel coding design targeted for the physical layer, if the Reed-Solomon (RS) code block has deletion beyond correction capability, the corresponding RS information block (FIG. 1) is a K turbo information packet 112 (FIG. 1). It will invariably be passed to the application layer 218, i.e., the source decoder, with a joint notification specifying which of the compromises has been compromised. Beyond the GF 256, the systemic structure of the external (N, K) RS code is not compromised, ie it allows direct use of correctly received turbo information packets.

When the erasure decoding fails due to a code block, i.e., the total number of deletions> (NK), the Reed-Solomon decoding layer retrieves the information block, along with information about the place of deletion or equivalent information, such as potentially corrupted byte locations. Still pass to an application layer 218, such as a video decoding layer.

Typical video signals are well known to be time-varying random signals, such as having varying space-time statistical properties. However, since this change is in most cases neither extremely fast nor structureless, there are significant spatial and temporal correlations both spatially and temporally that enable the large compression rates achieved by current video compression standards / algorithms. do.

The transformed / processed quantities and objects generated by the video compression algorithms from the pixel values of the raw digital video signal and the raw pixel values are both random variables / vectors. As such, in the range of video signals under consideration, these random variables / vectors have "a priori" probability distributions and other higher-dimensional statistics that characterize them in the absence of any additional information.

This spatiotemporal correlation manages not only the final quantity and objects (symbols) but also a number of intermediate products, which enter into the bitstream, the compressed representation of the raw digital video source. This means that the availability of any accurate / trusted information for any of these symbols makes it possible to reduce the uncertainty for any other symbol through its interaction with others. A very convenient mathematical model to explain this observation is through the use of posterior probability distributions. The posterior probability distribution reflects a change in the probabilistic description of these random symbols derived by indirect information obtained through correlated observations.

In one example, a context adaptable variable length code (CAVLC) coded syntax element is used in H.264. The context represents a classification of decoding states based on reliable information already received. Therefore, the context actually corresponds to the individual post-distribution of the symbols under consideration. In H.264, a number of symbols are encoded in the context, and the probability of each symbol varies with respect to the context in which it was coded, i.e., processed before the symbol. Many video compression standards, including H.264, implicitly or explicitly define and use several contexts, for example, the encoding type selected for a frame, I, P or B (where I, P and B, relatively, Intra, Predicted and Bi-directional) set the context that makes it possible for any macroblock type to be more generated while coding the frame. This requires a more efficient representation for this more likely macroblock type. Every context introduces the probability of collectively describing the state and / or framework into which the encoding / decoding process goes. This probability is determined after performing a number of tests, using different quality settings for different test signals, such as video sequences. While the signal is being encoded or decoded, this context changes and adapts to the actual input value.

The CAVLC coded symbol named coeff_token is used for illustrative purposes when coeff_token is a pseudonym of a vector quantity with two components. These are:

1. TotalCoeff ∈ {0, 1, 2, ..., 16}: the total number of nonzero transform coefficient levels in the 4 × 4 transformed and quantized block; And

2. TrailingOnes ∈ {0, 1, 2, 3}: Number of tracing '| value | = 1' transform coefficient levels in the same 4 x 4 block.

Analyzing and decoding coeff_token reveals these two positive values. The encoding and decoding process with variable length code (VLC) refers to the VLC code table. This code table has two entries, one of the original data, the possible symbol values, and the other of the corresponding codewords. In one form of entropy coding, loaded variable length coding or Huffman coding, less probable values are combined with / with longer codewords, while more probable values are combined with / with shorter codewords. Thus, by knowing the statistics of events / symbols for coding, if such statistics are sufficiently representative, it is possible to code such events / symbols with an average number of bits lower / less than required by fixed length coding. Table 1 reflects a partial table specifying five possible different contexts (each corresponding to one of the last five rows of the table) to decode coeff_token, which is just a VLC codebook that follows the five combined contexts. Rather, it is defined as a function of a variable named nC (each of which is fitted with a corresponding different post-distribution for the vector (TotalCoeff, TrailingOnes)).

Table 1

3 is an illustration of an example of a neighboring 4 x 4 block that provides correlated reference data for a corrupted current 4 x 4 block. In a simplified generalization, the context for determining the variable nC is specified as a function of the information available from neighboring 4 × 4 blocks A 302 and B 304. The correlation information from neighboring A 302 and B 304 determines the context for decoding the coeff_token in the corresponding binary representation, i.e., the current 4x4 block 306 in which the codeword of the bitstream is corrupted or partially corrupted. To help.

nA and nB are the number of nonzero transform coefficient levels in block A 302 located to the left of the current 4 x 4 block 306 and block B 304 located above the current 4 x 4 block 306, respectively (TotalCoeff (given by coeff_token). When both block A 302 and block B 304 are available, nC can be given by the following equation:

nC = (nA + nB + 1) >> 1. [One]

This equation simply describes nC as the arithmetic mean of nA and nB, always rounding the fractional part of the result to 0.5. This is one way that reliable correlated information available from neighboring 4 x 4 blocks A and B changes the decoder's expectation for information related to the current 4 x 4 block. This occurs here through the selection of context (which affects bitstream interpretation and analysis) based on the value of nC summarizing the posterior information available in any way. As can easily be determined from different lengths of codewords representing the same coeff_token value in different contexts, the context modifies the coeff_token prior probability distribution in accordance with the post information. It is an attribute of this embodiment that this post refined symbol probability distribution is used as a pre symbol probability function of the general MAP framework.

There are other less conventional methods that can be used to refine already received reliable reference data, i.e. post information, symbol probability distribution. If we hypothesize nA = 4, nB = 1, the 4 x 4 neighboring block B 304 does not vertically neighbor the slice boundary, but instead draws image details such as a vertical direct structure, e.g. vertical edge 308. Including, it is estimated based on its internal 4 x 4 prediction mode. The application of equation [1] gives nC = 3, which means a different context than either of neighbors A 302 and B 304. Under this situation, it is assumed that the codeword from the VLC codebook which certainly belongs to the context '2 ≦ nC <4' was used to encode coeff_token for the current block 306.

In another example, due to the statistically very frequent and therefore highly consistent presence of the vertical edge 308 in block B 304, similar vertical edge 310 is currently 4 × 4 block 306 from block B 304. It can be strongly expected to propagate. Thus, the coefficient structure of the current 4x4 block 306 and that of the neighboring 4x4 block B 304 are expected to be very similar to each other. Thus, in the current 4 × 4 block, the post refined probability distribution of coeff_token should be close to the one combined with the context of the neighboring block B 304. Thus, in this case for the current block 306, the codeword representing coeff_token will come from the '2 ≦ nC <4' context (as inferred by applying equation [1] above), and its post refinement. The probability distribution will be represented more accurately by the context '0 ≤ nC <2'.

As mentioned above, there is a significant difference between the following two probability distributions for coeff_token: (1) a dictionary PDF in its true real sense, ie a PDF without context or any other additional information available, and (2) neighbors If the 4 × 4 blocks A 302 and B 304 have non-zero coefficient level sums nA and nB respectively, then they are post-purified PDF, in some ways conditional PDF.

In one example, the statistical approach integrates a Bayesian framework for determining parameters, ie symbol values, maximum posterior probability (MAP) for evaluation. This Bayesian framework requires not only the conditional probabilistic model for a given parameter to be evaluated with relevant observations, but also the use of prior probabilistic models for the parameter. When the application layer (at the source decoder) 218 (upper FIG. 2) attempts to analyze and entropy decode the symbol 'coeff_token', the relevant observations and pieces of information for the coeff_token that it has access to at that point are: It can be classified into main classes:

1. no abnormalities already received and successfully decoded, ie reliable correlated information, especially those for neighbors;

2. A potentially corrupted (with suspect bit) bitstream segment carrying a codeword representing the value of coeff_token.

The Class 1 information is very important and in one example of a known method, the Class 1 type of post information is a Bayesian formulation by substituting a pure dictionary PDF of coeff_token by (post refined) PDF according to the appropriate context. Is integrated into. The described Bayesian model enables the use of all available information from the source (class 1 above) and the channel (class 2 above combined with the probabilistic channel model), hence the name joint source-channel entropy decoding. Let's do it.

4 is a block diagram of a joint source-channel MAP decoder for entropy decoding. At the application layer, joint source-channel entropy decoder 400 includes channel probability model generator component 402, memory module 1 404, MAP formulation and solution component 406, correlation data processing component 408, and memory 2. Module 410. In one embodiment, MAP formulation and solution component 402 may be realized through general purpose or application software running on an embedded CPU. All data 412 from the connected turbo / lead-solomon decoder 414, i.e. correctly received (including incorrectly received but then corrected) and incorrectly received symbols (deleted), are based on channel observations. Is sent to the channel probability model generator component 402 for developing the statistical properties of the channel. In one possible embodiment, the channel probability model generator may select and adopt one of a fixed set of probable channel models based on the most recent observation of the channel condition. The same data 412 is sent to the device 416 which acts as a switch to route the data from the MAC layer to the appropriate destination. From device 416, a correctly received symbol 418 is given a correlation data processing component that determines the post refined probability distribution function for an incorrectly received symbol 418, given all available and correlated reference information. 408). The correctly received symbol 418 is also sent to the video decoder 420 for regular decoding and display. An incorrectly received symbol 422 that cannot be corrected is sent to the MAP formulation and solution component 406 for possible error correction. The MAP formulation and solution component 406 receives data from the data routing switch device 416 and the memory modules 1 404 and 2 410 to calculate the maximum post probability function. The MAP function is used to determine the most likely value for an incorrectly received symbol through the maximization process. The symbols 424 corrected by the MAP formulation and solution component 406 are returned to their respective positions in the stream of symbols 418 sent to the video decoder 420.

Given the available knowledge obtained after performing one set of associated observations, the post-distribution is a synonym for the conditional probability distribution of symbols / parameters. To use this knowledge, the posterior distribution is marginalized through the use of Bayes' Rule and observations are conditioned by the hypothesis about the symbols / parameters to be evaluated. In the decoding method using the maximum post probability estimation framework, the following post PDF maximization problem needs to be solved:

[2]

[2]

Given a return of the value of Θ that achieves the observation x and the maximum, this will maximize the posterior probability of the unknown symbol Θ to be evaluated. In one embodiment, the observation x will be binary data, i.e. the bits present in the corrupted turbo information packet, and the randomness of x will result from the randomness of the channel which needs to be characterized by the effect of the appropriate probabilistic model. In equation [2] above, p (Θ | x) denotes the conditional probability of the symbol Θ given the observation x. Thus, it reflects the altered state of knowledge of Θ after observation, and thus reflects the name 'post probability'. The symbol Θ represents a bitstream symbol, for example, a macroblock type, a quantization parameter, an internal prediction direction, a motion vector (differential), a DC coefficient (differential), an AC coefficient, or the like. By BA's theorem, the maximization problem equivalent to that described in equation [2] is:

[3]

[3]

In one example, equation [3] is used to determine the actual value of an incorrectly received symbol. The term p (Θ) is the prior PDF (or prior probability) of the bitstream symbol Θ to be (soft) decoded. Hypotheses about Θ, i.e., the values allowed for Θ as well as the probabilities associated with these hypotheses (in the form of pre-continuous or discontinuous PDFs) are constructed on the basis of "no abnormal", ie reliable, correlated source observations. Will be. In one example, reliable correlation source observations are provided by the information correctly received at spatial and temporal neighboring entities such as, for example, neighboring blocks, macroblocks, slices and frames. Accurately received bits on either side of the known data, for example, a suspected bitstream segment, present a strict constraint that cannot be incompatible with any hypothesis. Also, no hypothesis can lead to syntax or semantic violations within a known data segment. All of these conditions are individually necessary for the feasibility of any hypothesis and, therefore, failure of one of them is sufficient to discard the hypothesis.

BS's rule replaces the post-PDF p (Θ | x) of the equation [2] with the result of the prior PDF p (Θ) and p (x | Θ) of the symbol Θ to obtain an equivalent equation [3]. do. Equivalence here is in terms of the result of the maximization problem. Given the symbol Θ, the function (conditional density) p (Θ | x) defines the probability of the observation x and thus, it is a function of the channel only and thus p (x | Θ) is a suitable (fixed or variable) Can be fully characterized by the channel model. On the other hand, by an important concept of the disclosed method, the dictionary PDF p (Θ) of the symbol Θ is replaced with a post-purified PDF for Θ based on the reliable correlation information available.

5 is a description of a temporally oriented frame. Abnormal source observations come from spatially neighboring blocks as shown by adjacent blocks A 302 and B 304 of FIG. 3 or from temporal relationships as shown in FIG. 5. As discussed above, the temporal change in the picture / frame content is such that a sudden change occurs between video frames 502 and 504 of the video sequence 506 that includes a set of consecutive and incorrectly received frames 508. May be slow to cause. Thus, a probabilistic model for lost symbols, such as providing a possibility for data suitable for corrupted region 510, can be obtained using uncorrelated reference data 512 and 514. This correlation data is used to provide post refinement to the prior probability distribution function of equation [3].

Referring to equation [3], the term p (x | Θ), which represents a probabilistic model (ie PDF) that conditionally describes observations from a given channel at the channel input, is a PDF derived from the measurement of the channel condition, It will be based on the analysis of the bit error pattern observed on the channel, namely the channel error model (channel probability model generator component 402 and memory module 1 404 of FIG. 4). Data for this purpose can be accumulated through actual channel simulations or from actual field tests where the actual transmitting and receiving subsystems are used with known input and recorded output signals.

The examples provided below illustrate the use of various probabilistic models. Any useful probabilistic model can be used to generate a PDF and is not limited to what will be described later. Table 2 describes the probability density function (PDF) of the residual burst length, ie, the streak length PDF for successive uncorrected Reed-Solomon (RS) codeword symbol deletion experienced by the mobile receiver. Table 2 uses (N, K) RS codes, ratio j / m turbo codes, and row-first scan patterns for data placement in RS information blocks.

Typical values for the various parameters described above and also mentioned in FIG. 1 may be described as follows:

N=16;

N = 16;

K ∈ {15, 14, 12};

K ∈ {15, 14, 12};

L=122; CRC 검사합계, 추적 터보 코딩 비트 (a.k.a 꼬리 비트) 및 보전 비트로 첨부된 추가적 3 바이트와 함께, 전체 터보 정보 패킷 크기는 125 바이트 = 1000 비트가 된다; 및

L = 122; The total turbo information packet size is 125 bytes = 1000 bits, with an additional 3 bytes appended with the CRC checksum, trace turbo coding bits (aka tail bits), and preservation bits; And

터보 코딩 비율 ∈ {1/3, 1/2, 2/3}

Turbo coding ratio ∈ {1/3, 1/2, 2/3}

Table 2. Sample Residual Burst Length Distribution

0.2) 보다 거의 2 배 가능하다 (확률

0.4). 이러한 통계는 오직 예로서만 이용할 수 있고 실제 실행에서 관측된 실제 통계가 다른 팩터뿐만 아니라 내부 및 외부 코드 비율에 강하게 따른다는 것을 주의해야 한다.Under this hypothesis, in the worst case, four consecutive turbo information packets are deleted. With reference to FIG. 1, since the application layer data, eg, video coding layer data, is scanned into a Reed-Solomon code block, the data from the worst case, which is a row-first, four consecutive deleted turbo information packets, is an RS codeword. In each row of FIG. 1 defining s i will be mapped to 4 contiguous GF 256 symbols, ie 4 bytes. Under this hypothesis, in the worst case example, the decoder decodes through a maximum that is 4 bytes (32 bits) of application layer data corrupted by bit errors in unknown places. According to these sample statistics, a burst of up to 4 possible lengths has a probability of some other burst length (1,2 or 3 respectively).

Almost twice as likely (0.2)

0.4). It should be noted that these statistics are available only as examples and that the actual statistics observed in actual execution strongly depend on the internal and external code rates as well as other factors.

Table 3 shows a PDF of the total number of bits of error in the deleted turbo information packet. This statistic can be used to infer the average number of bit errors for the dropped packet or the prior probability of the bits (in the dropped packet) present in the error.

Table 3. Distribution of total number of bit errors in dropped turbo information packets

Table 4 shows a PDF of the variance of the bits of the error in the deleted turbo information packet. This is just the error burst length in the packet. The sample statistics in Table 4 typically indicate that the bits of error are not concentrated in one local but rather span the entire packet.

Table 4. Distribution of Error Burst Length in Dropped Turbo Information Packets

Table 5 describes a PDF of the bitstream length of the error, ie, the sum of consecutive bits that are all errors, in the deleted turbo information packet. As it becomes clear, this particular statistic is very closely related to the bit error pattern and plays a very important role in the examples discussed below.

Table 5. Distribution of Number of Consecutive Bits of Error in Dropped Turbo Information Packets

Finally, Table 6 reflects the PDF of the distance separating the neighboring streaks of bit error in the deleted turbo information packet. Like the statistics reported in Table 5, this particular statistic is also very closely related to a simple and efficient model for bit error patterns.

Table 6. Distribution of Distances Between Bit Error Streaks in Dropped Turbo Information Packets

In one example, the probabilistic models of Tables 7, 8 and 9 are used below. Without loss of generality, and for illustrative purposes only, it is assumed that the decoder analyzes the coeff_token syntax element from the bitstream and only entropy decoding, ie, continuously. There are two reasons for this hypothesis. The first is that this setting provides a suitable example to illustrate the basic idea, without the need of introducing a codebook of other syntax elements that would complicate the described example. Second, there is no immediate approach to PDF of other syntax elements that requires similar encoding simulation-based statistical analysis studies to be performed for PDFs of other syntax elements.

Table 7. Sample post-purified PDF of coeff_token for context '0 <nC <2' (from Table 1)

For coeff_token's dictionary PDF, a post-purified PDF that follows its basic context is used for the hypothetical basic context. As noted previously, for example, stronger expected correlation to the current block (block 306 of FIG. 3) (vertical edge 308 likely to continue as vertical edge 310 of FIG. 3). Through the identification of a neighboring block (block B 304 of FIG. 3) with an even context based post-purified PDF can be modified to reflect this estimate.

Table 8. Sampling of streak length PDFs for consecutive bit errors, from Table 5

Table 9. Sample truncated, ie, partial PDF of the inter-error distance characterized as the number of bits, from Table 6

' 및 '?' 는 각각 이상없는 비트와 잠재적으로 훼손된 비트를 표시하기 위해 사용되고 'X' 는 이 예에 관계가 없는 값의 비트를 지시한다. 비트는 또한 그들의 위치 지표에 의해 참조되고 비트 위치 넘버링은 좌측에서 우측으로 1 로 시작하여 24 까지 올라가며, 비트 25 및 26 이 또한 도시됨을 주의한다.In one example, a decoder is provided with the next three bytes as part of the bitstream received by the physical layer. The FEC layer has already processed the data and marked the middle byte while leaving the dropped packet. Therefore, all bits of the middle byte are suspected of being corrupted by bit errors. In the following,

'And'? ' Are used to indicate bits that are both intact and potentially corrupted, respectively, and 'X' indicates bits of values that are irrelevant to this example. Note that the bits are also referenced by their position indicators and the bit position numbering starts from 1 from left to right and goes up to 24, and bits 25 and 26 are also shown.

In this example, the decoder successfully decoded coeff_token (TrailingOnes, TotalCoeff) = (1, 2) with the codeword '000100' including bit positions 2 through 7. As discussed above, the intact bits present on one side of a potentially corrupted byte establish important constraints to prove / evaluate various hypotheses that a decoder may generate. Additional constraints include:

이용 가능한 이웃들의 블록 경계를 가로지른 연속성과 같은 결과적인 픽셀 도메인 형상의 연속성/평탄성; 및

Continuity / flatness of the resulting pixel domain shape, such as continuity across the block boundaries of available neighbors; And

의심되는 오류 로케이션 (location) 뒤로 연속된 이상없는 비트의 어떤 수에 대한 분석 및 엔트로피 디코딩 실패가 없는 것, 즉 신택스 및 시맨틱 체크 실패가 없는 것;

No analysis and entropy decoding failures for any number of consecutive noncontiguous bits behind a suspected error location, ie no syntax and semantic check failures;

The use of is important and recommended because they will enhance the success of the proposed algorithm. Both of these constraints are used to increase the level of confidence combined with a feasible hypothesis. For example, a particular hypothesis about the original value of a corrupted symbol may lead to a relatively high posterior probability and is consistent with the first few leading bits of the first prime number immediately following the potentially corrupted bitstream segment in which it exists. However, if subsequent decoding based on this hypothesis leads to subsequent syntax or semantic failure in the bitstream in the range of bitstreams known to have been correctly received, this is a sufficient indication that the underlying hypothesis was not correct and should be discarded. As such, each and every symbol correctly decoded syntactically and semantic following the selection of a particular hypothesis actually increases the level of confidence of that hypothesis. Details in the implementation of this idea will vary from embodiment to embodiment.

Starting at bit position 8, the following sequence of coeff_token values can be decoded from the bitstream: (1, 1) (1, 1) (2, 3) [(0, 3) or (1, 4)] → '01' '01' '0000101' '00000011X'.

From Table 7, the non-memory, i.e., a priori probability hypotheses and combined based on the independent symbols, model (more precisely, the post refined probability measure) is 0.2999 × 0.2999 × 0.0077 ~ = 6.9254 × 10 - a ⁴ (P ((0, 3 )) Or P ((1, 4)) were not included in this calculation because all of the corresponding codewords are entirely within the range of bits that are abnormal.

The prior probabilities for this hypothesis should be scaled by the conditional probabilities of observations from channels in this hypothesis. This hypothesis means that there is no bit error in the suspect intermediate byte. In this case, given every 8 bits, either the expected value of 2 bits of error (~ 250 bit error on average among 1000 bits) or 8 bits ≤ probability of inter-streak distance between implied (implied) errors is given. The absence of any of the suspected bits is an unlikely event and therefore corresponds to a fairly small probability. Therefore, P (inter-error-streak distance ≥ 8) = 0.095 can be used as the conditional probability of observation from the channel in this hypothesis. With this scaling, the probability measure coupled with the hypothesis, i.e., P ('000000 10100001 0100000011 ...' transmitted and received without any bit error) is 6.5791 × 10 ⁻⁵ .

In this example, since there are only 8 bits potentially corrupted by bit error, the actual error pattern will be one in 256 possible 8-bit long binary sequences. In this representation, 0 means 'error' bit position and 1 means 'error' bit position. Of these 256 possible error patterns (error masks), some become ineligible from immediately becoming potential hypotheses. For example, such a pattern comprising the subsequence '... 101 ...' would have an overall probability of zero (from Table 9) since it would correspond to an inter-error distance of 1 with a probability of zero (from Table 9). Will be allocated. Thus, they can be immediately excluded from the list of potential error pattern hypotheses of the decoder. Although the decoder may not shorten this list as described because of its different error pattern statistics, it may be more or less unsuitable for patterns that include excessively or excessively small bit errors in terms of the expected rate of bit errors. Evaluation of any corresponding error pattern hypothesis can still be excluded, postponed, or conditionally started.

As an executable hypothesis, consider the error pattern '10010001' as an example in which the decoder checks / evaluates one possible error pattern. This error pattern effectively removes the error from the received channel output and, when applied to the intermediate byte, i.e. ex-or'ed, with the next three byte sequence (again with additional 25th and 26th bits) ) Reaches:

X 0 0 0 1 0 0 0 0 0 1 1 0 0 0 0 0 1 0 0 0 0 0 0 1 1 X.

Starting at bit position 8, the following sequence of coeff_token values can be decoded from the bitstream:

(3, 3) (3, 6) (3, 4) → '00011' '00000100' '000011'

From Table 7, our previous nonmemory, that is, independent symbols, prior probabilities (exactly post-purified probability measures) combined with hypotheses based on models are:

0.0226 × 0.0020 ~ = 4.52 × 10 - 5 And P ((3, 4)) was not included in this calculation because the corresponding codeword is entirely within the range of bits that are abnormal.

As before, the prior probabilities for this hypothesis should be scaled by the conditional probabilities of observations from channels in this hypothesis. The decoder may circumvent this calculation, noting that the probability is already less than the probability measure (6.5791 × 10 ⁻⁵ ) combined with the 'bit error free' hypothesis. However, this calculation continues to be completed and to provide another example of this process. The hypothesized error pattern is '10010001'. This can be broken down into the following sequence of events in the channel:

'One bit error'; 'Between two errors-streak distance'; 'One bit error'; 'Between three errors-streak distance'; 'One bit error'.

Again hypothesizing a simple non-memory model of error pattern, the probability combined with the error pattern can be described as follows:

0.7 × 0.3 × 0.7 × 0.22 × 0.7 to = 0.0226.

With this scaling, the probability measure combined with the current hypothesis, namely P ('000000 00110000 01000000 11 ...' sent and received with error pattern '10010001' that corrupts the middle byte) is 1.0232 × 10 ^{− 6} becomes

Only based on the comparison of these two posterior possibilities, the decoder would prefer the first hypothesis:

'' Sent and received without bit error '... 0001000 10100001 01000000 11 ...' ''.

Of course, the decoder needs to evaluate all possible hypotheses in the general framework described above in some way and, more importantly, need to include additional constraints in this assessment to improve the error correction rate. Although the above example considered the corresponding probabilities and hypotheses based on the decoding of one symbol as an example in which coeff_token was used, a similar method could be applied to a vector of symbols containing one or more symbols. Such symbols may be of the same type, ie homogeneous vectors, or may be of different types, ie heterogeneous vectors. Considering vectors of symbols, ie complex symbols instead of one symbol at a time, results in a more efficient and better structured formulation that improves more constraints, additional statistical information, and error correction rates and lowers computational complexity. Can provide opportunities for

In regular bitstream syntax, symbols with varying amounts of materiality for the restructured signal are interleaved together. Some slice data partitioning (SDP) makes it possible to classify coded representations of syntax elements with similar / comparative significance. This reshaping has the following distinct advantages. If a less significant partition, i.e. a chunk of data, is corrupted by error, it can be simply and safely ignored, i.e. excluded, and this corruption does not corrupt the data of other partitions that can be used to achieve lower quality but still useful restructuring. It is not. Something like SDP can be used to promote an increase in efficiency as well as a reduction in computational complexity combined with other embodiments.

In one example, the disclosed method is conditionally turned on and executed only when high materiality SDP splitting is compromised. Otherwise, i.e., in the case of a lower materiality partition, the data of the corrupted partition may only be used in part, i.e. only the parts that are known to be good are used, or both to avoid the additional complexity required to process it. May be excluded.

In another example of the disclosed method used in combination with SDP, the computational efficiency of the error correction scheme is improved. This is useful in an embedded environment characterized by computational, memory or power limitations. The disclosed method can request access to a significant amount of data, such as probability models for many different types of symbols. Cache efficiency of embedded devices is an important factor in determining execution speed and processor overhead. Not requesting a very large set of data in a short execution period can result in the avoidance of cache data trashing or frequent cache rewriting, thus resulting in an increased cache hit ratio. SDP reduces the type diversity of syntax elements in each data segment. Instead, this increases the coherence of the data required to process a limited subset of syntax elements. This reduction in the number of syntax elements may also result in a corresponding reduction in the amount of reference data needed to assign probabilities to the error pattern hypothesis.

6 is an example of a flow diagram for a method of MAP based joint source-channel entropy decoding. The coded and modulated symbols in the bitstream are received and demodulated at the demodulator of the wireless remote device (step 602). Receiving means, such as demodulator 210 of FIG. 2, may perform step 602. A symbol can provide information about any type of compressed data, including but not limited to compressed multimedia data such as macroblock types, quantization parameters, internal prediction directions, motion vectors, DC coefficients, or AC coefficients. After demodulation, the symbol is passed through an error correction decoder such as a connected turbo / lead-solomon decoder. (Step 604). Decoding means and correction bit determining means, such as the turbo / lead-solomon decoder 200 of FIG. 2, may perform step 604. After error detection and correction at the Turbo / Lead-Solomon decoder, the device acting as a switch (see component 416 of FIG. 4) may have corrupted or partially corrupted symbols in the MAP formulation and solution components of the application layer. (See component 406 of FIG. 4). The output from the turbo / lead-solomon decoder is also sent to a channel probability model generator (see component 402 of FIG. 4) for determining, ie calculating or selecting, the appropriate probable characteristic of the current channel state. The data routing switch sends the correlation data to the correlation data analysis component (see component 408 of FIG. 4). Correlation data may be obtained, for example, from a symbol located near a spatially or temporally corrupted symbol (step 610). The MAP formulation and solution component accepts data from memory modules 1 and 2 (see modules 404 and 410 of FIG. 4) and also symbols that cannot be corrected from the data routing switch (step 612), for corrupted symbols. Resolve the maximize problem that maximizes post PDF (step 614). Thus, the MAP formulation and solution component may calculate the most probable value for a symbol that is corrupted or partially corrupted (ie, incorrectly received and corrected) (step 614). Hypothesis determining means and probability allocation means, such as the MAP formulation and solution component 406 of FIG. 4, may perform step 614. The corrected symbols from the MAP formulation and solution components are combined with the correctly received (including incorrectly received but subsequently corrected by FEC) symbols (step 616), and the data is sent to the video decoder, The output is sent to the display unit (step 618).

The disclosed method can be applied not only to H.264 bitstreams but also to processes of other currently available or future data compression schemes designed for video, image, audio, and other forms of media. By way of example, the present invention can be applied to Context-Based Adaptive Arithmetic Coding (CABAC). The discussed MAP optimization approach can be adapted to the 'CABAC' arithmetic coding scheme because CABAC uses similar probability modeling to generate context for symbols that are “binaryized” or converted to binary code. In CABAC, non-binary-value symbols (eg, transform coefficients or motion vectors) are binarized before arithmetic coding. This process is similar to the process of converting a symbol to a variable length code, but the binary code is also encoded (by an arithmetic coder) before transmission. A "context model" is a probability model for one or more bits of a binary symbol. This model may be selected from a selection of available models that follow the statistics of the recently-coded symbols. The context model stores the probability of each bit becoming either "1" or "0". This selected context model is then updated based on the actual coded value. In accordance with the conditional PDF describing the channel condition, as described above, in order to refine the raw pre-symbol distribution, the application of post-information, i.e., correlation data from neighboring (spatial / temporal) symbols, is the CABAC symbol of the error. Formulate a MAP optimization approach for the entropy decoding of.

Aspects of the disclosed examples include, but are not limited to, the following description.

Receiving a bitstream comprising one or more bits, determining whether the bitstream has one or more corrupted bits, determining one or more hypotheses representing an error pattern, and assigning probabilities to each of the hypotheses And the probability is determined based on one or more reference data.

Means for receiving a bitstream comprising one or more bits, means for determining whether the bitstream has one or more corrupted bits, means for determining one or more hypotheses representing an error pattern, and probability for each of the hypotheses Means for assigning a value, wherein the probability is determined based on one or more reference data.

Error correction configured to receive a bitstream comprising one or more bits, determine whether the bitstream has one or more corrupted bits, determine one or more hypotheses that represent error patterns, and assign probabilities to each of the hypotheses An electronic device, wherein the probability is determined based on one or more reference data.

Receiving a bitstream comprising one or more bits, determining whether the bitstream has one or more corrupted bits, determining one or more hypotheses representing an error pattern, and assigning probabilities to each of the hypotheses And means for causing a computer to execute a method wherein the probability is determined based on one or more reference data.

In addition, those skilled in the art will appreciate that the various illustrative logical blocks, modules described in connection with the embodiments disclosed herein, and the algorithm steps described in connection with the examples disclosed herein may be electronic hardware, computer software, or combinations thereof. It can be seen that it can also be implemented. To clearly illustrate this alternative possibility of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above primarily in terms of their functionality. Whether such functionality is implemented in hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but should not be construed as causing a determination of such implementation to depart from the scope of the present invention.

The various illustrative logic blocks, modules, circuits described in connection with the embodiments disclosed herein may be a general purpose processor, digital signal processor (DSP), application specific integrated circuit (ASIC), field programmable gate array signal (FPGA), Or other programmable logic device, separate gate or transistor logic, separate hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in other ways, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, eg, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other configuration.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, software module, or a combination of two executed by a processor. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, which can read information from and write information to the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside within an ASIC. The ASIC may reside in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal.

Thus, by performing an optimization problem formulation within the maximum posterior probability (MAP) framework and entropy coded, its representation is improved by solving it to determine the approximation of the hypothesis for the symbol, which is corrupted with bad bits. A method and apparatus for providing correction capability have been described.

Claims (34)

delete delete Receiving a bitstream comprising one or more bits; Determining if the bitstream has one or more corrupted bits; Determining one or more hypotheses representing the error pattern; Determining a prior probability distribution function based on channel condition measurements; And Assigning a probability to each said hypothesis, And the probability is determined based on one or more reference data, the reference data comprising a prior probability distribution function. Receiving a bitstream comprising one or more bits; Determining if the bitstream has one or more corrupted bits; Determining one or more hypotheses representing the error pattern; Decoding uncorrupted bits in one or more frames; And Assigning a probability to each said hypothesis, Wherein the probability is determined based on one or more reference data, the reference data comprising one or more of the decoded uncorrupted bits. Receiving a bitstream comprising one or more bits; Determining if the bitstream has one or more corrupted bits; Determining one or more hypotheses representing the error pattern; Decoding an uncorrupted bit in a frame comprising the corrupted bit; And Assigning a probability to each said hypothesis, Wherein the probability is determined based on one or more reference data, the reference data comprising one or more of the decoded uncorrupted bits. Receiving a bitstream comprising one or more bits; Determining if the bitstream has one or more corrupted bits; Determining one or more hypotheses representing the error pattern; Assigning probabilities to each of the hypotheses; Decoding the uncorrupted bits; Determining whether one of the hypotheses leads to a syntax or semantic failure within the uncompromised bit; And Assigning a zero probability to the hypothesis that leads to the failure, wherein the probability is determined based on one or more reference data. Receiving a bitstream comprising one or more bits; Determining if the bitstream has one or more corrupted bits; Determining one or more hypotheses representing the error pattern; And Assigning a probability to each said hypothesis, The probability is determined based on one or more reference data, Determining the hypothesis further comprises determining the hypothesis for a group of bits that make up a coded symbol. Receiving a bitstream comprising one or more bits; Determining if the bitstream has one or more corrupted bits; Determining one or more hypotheses representing the error pattern; And Assigning a probability to each said hypothesis, The probability is determined based on one or more reference data, Determining the hypothesis further comprises determining a hypothesis for the group of bits that make up the coded symbol, And the coded symbol is selected from the group consisting of macroblock type, quantization parameter, internal prediction direction, motion vector, DC coefficient and AC coefficient. Receiving a bitstream comprising one or more bits; Determining if the bitstream has one or more corrupted bits; Determining one or more hypotheses representing the error pattern; Decoding uncorrupted bits in one or more slices; And Assigning a probability to each said hypothesis, Wherein the probability is determined based on one or more reference data, the reference data comprising one or more of the decoded uncorrupted bits. Receiving a bitstream comprising one or more bits; Determining if the bitstream has one or more corrupted bits; Determining one or more hypotheses representing the error pattern; Inaccurate using information selected from the group consisting of bit-error-stre distance, error burst length, average number of bit errors for dropped packets, neighboring bytes without bit errors, user data fields, and error characteristics of received packets. Determining a probability distribution for the received symbol; And Assigning a probability to each said hypothesis, And the probability is determined based on one or more reference data, the reference data comprising the probability distribution. delete delete delete Means for receiving a bitstream comprising one or more bits; Means for determining if the bitstream has one or more corrupted bits; Means for determining one or more hypotheses representing the error pattern; Means for assigning a probability to each said hypothesis; And Means for determining a prior probability distribution function based on channel condition measurements; And the probability is determined based on one or more reference data, the reference data comprising the prior probability distribution function. Means for receiving a bitstream comprising one or more bits; Means for determining if the bitstream has one or more corrupted bits; Means for determining one or more hypotheses representing the error pattern; Means for decoding an uncorrupted bit in one or more frames; And Means for assigning a probability to each said hypothesis, And the probability is determined based on one or more reference data, wherein the reference data includes one or more of the decoded uncorrupted bits. Means for receiving a bitstream comprising one or more bits; Means for determining if the bitstream has one or more corrupted bits; Means for determining one or more hypotheses representing the error pattern; Means for assigning a probability to each said hypothesis; And Means for decoding an uncorrupted bit in a frame comprising the corrupted bit; The probability is determined based on one or more reference data, And the reference data includes one or more of the decoded uncorrupted bits. Means for receiving a bitstream comprising one or more bits; Means for determining if the bitstream has one or more corrupted bits; Means for determining one or more hypotheses representing the error pattern; Means for assigning a probability to each said hypothesis; Means for decoding an uncorrupted bit; Means for determining whether one of the hypotheses leads to a syntax or semantic failure within the uncompromised bit; And Means for assigning a zero probability to the hypothesis that leads to the failure, wherein the probability is determined based on one or more reference data. Means for receiving a bitstream comprising one or more bits; Means for determining if the bitstream has one or more corrupted bits; Means for determining one or more hypotheses representing the error pattern; And Means for assigning a probability to each said hypothesis, The probability is determined based on one or more reference data, And means for determining the hypothesis further comprises means for determining the hypothesis for a group of bits that make up a coded symbol. Means for receiving a bitstream comprising one or more bits; Means for determining if the bitstream has one or more corrupted bits; Means for determining one or more hypotheses representing the error pattern; And Means for assigning a probability to each said hypothesis, The probability is determined based on one or more reference data, Means for determining the hypothesis further comprises means for determining the hypothesis for the group of bits that make up the coded symbol, And the coded symbol is selected from the group consisting of macroblock type, quantization parameter, internal prediction direction, motion vector, DC coefficient and AC coefficient. Means for receiving a bitstream comprising one or more bits; Means for determining if the bitstream has one or more corrupted bits; Means for determining one or more hypotheses representing the error pattern; Means for assigning a probability to each said hypothesis; And Means for decoding an uncorrupted bit in one or more slices; The probability is determined based on one or more reference data, And the reference data includes one or more of the decoded uncorrupted bits. Means for receiving a bitstream comprising one or more bits; Means for determining if the bitstream has one or more corrupted bits; Means for determining one or more hypotheses representing the error pattern; Means for assigning a probability to each said hypothesis; And Inaccurate using information selected from the group consisting of bit-error-stre distance, error burst length, average number of bit errors for dropped packets, neighboring bytes without bit errors, user data fields, and error characteristics of received packets. Means for determining a probability distribution for the received symbol; And the probability is determined based on one or more reference data and the reference data includes the probability distribution. delete delete delete Receive a bitstream comprising one or more bits, determine whether the bitstream has one or more corrupted bits, determine one or more hypotheses that represent error patterns, assign probabilities to each of the hypotheses, and channel conditions Determine a prior probability distribution function based on the measurement, The probability is determined based on one or more reference data, And the reference data comprises the prior probability distribution function. Receive a bitstream comprising one or more bits, determine whether the bitstream has one or more corrupted bits, determine one or more hypotheses that represent an error pattern, assign probabilities to each of the hypotheses, and Configured to decode uncorrupted bits in a frame, The probability is determined based on one or more reference data, And the reference data comprises one or more of the decoded uncorrupted bits. Receive a bitstream comprising one or more bits, determine whether the bitstream has one or more corrupted bits, determine one or more hypotheses that represent error patterns, assign probabilities to each of the hypotheses, and Is configured to decode an uncorrupted bit in a frame comprising the bit, The probability is determined based on one or more reference data, And the reference data comprises one or more of the decoded uncorrupted bits. Receive a bitstream containing one or more bits, Determine if the bitstream has one or more corrupted bits, determine one or more hypotheses that represent error patterns, assign probabilities to each of the hypotheses, Decoding an uncorrupted bit, determining whether one of the hypotheses within the uncorrupted bit leads to a syntax or semantic failure, and Assign a probability of zero to the hypothesis that leads to the failure, And the probability is determined based on one or more reference data. Receive a bitstream comprising one or more bits, determine whether the bitstream has one or more corrupted bits, determine one or more hypotheses that represent error patterns, assign probabilities to each of the hypotheses, and code Determine the hypothesis for the group of bits that make up a symbol, And the probability is determined based on one or more reference data. Receive a bitstream comprising one or more bits, determine whether the bitstream has one or more corrupted bits, determine one or more hypotheses that represent error patterns, assign probabilities to each of the hypotheses, and code Determine the hypothesis for the group of bits that make up the symbol, The probability is determined based on one or more reference data, And the coded symbol is selected from the group consisting of macroblock type, quantization parameter, internal prediction direction, motion vector, DC coefficient and AC coefficient. Receive a bitstream containing one or more bits, determine whether the bitstream has one or more corrupted bits, determine one or more hypotheses that represent error patterns, assign probabilities to each of the hypotheses, and Configured to decode the uncorrupted bits in the slice, The probability is determined based on one or more reference data, And the reference data comprises one or more of the decoded uncorrupted bits. Receive a bitstream containing one or more bits, Determine if the bitstream has one or more corrupted bits, Determine one or more hypotheses that represent error patterns, Assign a probability to each of these hypotheses, and Bit-error-streak distance, error burst length, average number of bit errors for dropped packets, neighboring bytes without bit errors, user data fields, and errors in received packets Determine a probability distribution for an incorrectly received symbol using information selected from the group, The probability is determined based on one or more reference data, And the reference data comprises the probability distribution. delete delete

Images