JP6312312B2 - Context model generation device, encoding device, and decoding device - Google Patents

Description

  The present invention relates to a context model generation device, an encoding device, and a decoding device.

  For entropy coding in source coding, an arithmetic code that divides the number line into sections according to the occurrence probability of the symbol to be encoded, or a variable-length code word according to the occurrence probability of the symbol to be encoded is used. There is a Huffman code to assign. Such entropy coding enables more efficient coding than a fixed-length coding method.

  In some cases, the appearance tendency of data to be encoded can be predicted from data (hereinafter referred to as reference data) different from data to be encoded (hereinafter referred to as target data). In this case, when encoding, adaptively changing the occurrence probability model of the target data according to the reference data, the encoding can be performed more efficiently than the encoding method using the fixed occurrence probability model. Is possible. The tendency of appearance of the target data according to the reference data is called a context, and the mechanism for generating the occurrence probability model of the target data from the reference data is called a context model.

  AVC / H. H.264, HEVC / H. In H.265, context adaptive arithmetic coding called CABAC (Context-based Adaptive Binary Arithmetic Coding) is employed. In CABAC, the context is updated based on statistics of previous input symbols. A specific method for mounting the CABAC is disclosed in Patent Document 1.

  In addition to the context selection based on the statistics of the previous input symbol, a method for improving the coding efficiency by performing context selection using the current input symbol has also been proposed (for example, Patent Document 2).

Japanese Patent No. 4886755 Japanese Patent No. 5221047

  However, in the above encoding method, a context model is selected from a plurality of context models prepared in advance. For this reason, there is a problem that there may be no context model close to the context of the target data.

  The present invention has been made in view of such circumstances, and provides a context generation device, an encoding device, and a decoding device capable of providing a context closer to the context of target data.

  The present invention has been made to solve the above-described problem, and one aspect of the present invention provides a context that generates a context model that indicates the occurrence probability of each data value of target data when a reference data value is given. A model generation device, a prior probability estimation unit for calculating a first model composed of occurrence probabilities of data values in first learning data representing the target data, and second learning data representing the target data , A context estimation unit for calculating a second model composed of the occurrence probability of each data value when a reference data value is given, the first model, and the occurrence of each data value when a reference data value is given A mixture ratio determining unit that is a candidate model composed of probabilities and determines a mixture ratio between a predetermined candidate model and the first model with reference to the second model , And the said first model candidate model, a context model generating device characterized by comprising a context model generator for generating a context model mixed in the mixing ratio.

  Another aspect of the present invention is the context model generation device described above, wherein the context model generation unit linearly combines the first model and the candidate model with the mixing ratio and the context model. Generating.

  Another aspect of the present invention is the context model generation apparatus described above, wherein the candidate model is a Kronecker delta function.

  According to another aspect of the present invention, a prior probability estimation unit that calculates a first model composed of occurrence probabilities of data values in first learning data representing target data, and second learning representing the target data. Context estimation unit for calculating a second model composed of the occurrence probabilities of each data value when a reference data value is given, the first model, and each data value when the reference data value is given A mixture ratio determining unit that determines a mixture ratio between a predetermined candidate model and the first model with reference to the second model, and the first model. Using a context model generation unit that generates a context model in which the candidate models are mixed at the mixing ratio, and a context model generated by the context model generation unit By encoding the target data, an encoding device, characterized in that it comprises an encoding unit that generates encoded data.

  According to another aspect of the present invention, there is provided the above-described encoding device, a multiplexing unit that multiplexes encoded data generated by the encoding unit and information indicating a mixing ratio determined by the mixing ratio determining unit It is characterized by providing.

  According to another aspect of the present invention, a prior probability estimation unit that calculates a first model composed of occurrence probabilities of data values in first learning data representing target data, and second learning representing the target data. Context estimation unit for calculating a second model composed of the occurrence probabilities of each data value when a reference data value is given, the first model, and each data value when the reference data value is given A mixture ratio determining unit that determines a mixture ratio between a predetermined candidate model and the first model with reference to the second model, and the first model. A context model generation unit that generates a context model in which the candidate model is mixed at the mixing ratio, and encoded data in which the target data is encoded is converted into the context model. A decoding apparatus characterized by using a context model generator has generated and a decoder for decoding.

  According to another aspect of the present invention, a separation unit that separates input data into encoded data obtained by encoding target data and information indicating a mixing ratio used for encoding the encoded data; Prior probability estimation unit for calculating a first model composed of the occurrence probability of each data value in the first learning data representing the target data, the first model, and each data value when a reference data value is given A context model generation unit that generates a context model obtained by mixing a predetermined candidate model with the mixture ratio, and encoded data obtained by encoding the target data. And a decoding unit that performs decoding using the context model generated by the context model generation unit.

  According to the present invention, a context closer to the context of the target data can be provided.

It is a schematic block diagram which shows the structure of the information source encoding / decoding system 10 by 1st Embodiment of this invention. It is a schematic block diagram which shows the structure of the encoding apparatus 100 in the embodiment. It is a schematic block diagram which shows the structure of the decoding apparatus 200 in the embodiment. It is a schematic block diagram which shows the structure of the encoding apparatus 100a by 2nd Embodiment of this invention. It is a schematic block diagram which shows the structure of the decoding apparatus 200a in the same embodiment. It is a table which shows the example (the 1) of LUT which defines the candidate model Q in the modification of each embodiment. It is a table which shows the example (the 2) of LUT which defines the candidate model Q in the modification of each embodiment.

(First embodiment)
Hereinafter, a first embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a schematic block diagram showing the configuration of an information source coding / decoding system 10 according to the first embodiment of the present invention. The information source encoding / decoding system 10 includes an encoding device 100 and a decoding device 200. The encoding apparatus 100 generates encoded data E by entropy encoding the target data S. The encoding device 100 generates and uses a context model when performing entropy encoding. The decoding device 200 decodes the encoded data E, and generates decoded data D obtained by restoring the information source S. The encoded data E may be input to the decoding device 200 via a communication path or the like, or input to the decoding device 200 from a recording medium such as a DVD (Digital Versatile Disc) or an HDD (Hard Disk Drive). May be.

  FIG. 2 is a schematic block diagram showing the configuration of the encoding apparatus 100. The encoding apparatus 100 includes an occurrence probability generation apparatus 111, a buffer unit 101, and an entropy encoding unit 108. The occurrence probability generation device 111 includes a context model generation device 110, a reference data acquisition unit 106, and an occurrence probability model selection unit 107. The context model generation device 110 includes a prior probability estimation unit 102, a context estimation unit 103, a mixing coefficient determination unit 104, and a context model generation unit 105.

The buffer unit 101 accumulates the target data S. The target data S accumulated in the buffer unit 101 is used as learning data for generating a context model. The prior probability estimation unit 102 estimates the prior probability of X in the target data S from the target data S accumulated in the buffer unit 101. In the present embodiment, X is the target data S divided into 4 bits. Specifically, the prior probability estimation unit 102 reads the target data S stored in the buffer unit 101 as learning data L1. Then, the prior probability estimation unit 102 sets Expression (1) as the prior probability Pr (X), where the occurrence probability of each value that X can take in the sequence X _n obtained by dividing the read learning data L1 into 4 bits each. Use to calculate.

The context estimation unit 103 reads the target data S stored in the buffer unit 101 as learning data L2, and estimates the context in the target data S from the learning data L2. Specifically, the context estimation unit 103 uses N values X _n (nε {1, 2,..., N}) from the buffer unit 101 and reference data for each value X _n as the learning data L2. Read Y _n (n∈ {1, 2,..., N}). The context estimation unit 103 calculates the context Pr (X | Y) in the read learning data L2 using Expression (2).

In this embodiment, _Xn is obtained by dividing the target data S by 4 bits. Y _n is the 4 bits immediately before X _n in the target data S, respectively. For example, _{Y 3} is a 4-bit immediately before the _{X 3.} However, the bit length of _Xn may be shorter or longer than 4 bits. Also, the bit length of Y _n may be less than 4 bits may be long, may be different from the X _n.

Further, Y _n, for example, such as 10-bit previous X _n, may not be immediately before the X _n, furthermore, X _n may be information such as what number in the target data S. For example, if the target data S is a parameter string arranged in the order of raster scans related to an image, if Y _n is immediately before X _n , Y _n is a parameter related to the left adjacent portion of the target portion of X _n. . However, when the Y _n and one line before in the raster scan of X _n, Y _n is the X _n is a parameter related to neighboring portion on the portion of interest.

  The mixing coefficient determination unit 104 stores in advance a candidate model Q (X | Y) including the occurrence probability of each data value X when the reference data value Y is given. The mixing coefficient determination unit 104 determines a mixing coefficient between the candidate model Q (X | Y) and the prior probability Pr (X) estimated by the prior probability estimation unit 102. At this time, the mixing coefficient determination unit 104 determines that the model P (X | Y) resulting from the linear combination of the models using the mixing coefficient a, that is, the mixing represented by Expression (3), is the context estimation unit 103. Is determined as close as possible to the estimated context Pr (X | Y).

  Specifically, the mixing coefficient determination unit 104 calculates a hat that becomes 0 when the partial differentiation of the expression (4) with respect to a is performed by the expression (5) and sets it as the mixing coefficient.

In equations (4) and (5), (X, Y) given to sigma indicates that only the combination of X and Y included in the learning data L2 is to be summed. In this way, the context Pr (X | Y) in the learning data L2 need not be obtained for all combinations of X and Y.
In the present embodiment, as K (X | Y), Kronecker delta δ _{X, Y} having arguments X and Y shown in Expression (6) is used.

  The context model generation unit 105 uses the mixing coefficient a hat determined by the mixing coefficient determination unit 104 for the prior probability Pr (X) estimated by the prior probability estimation unit 102 and the candidate model Q (X | Y). A mixed context model P (X | Y) is generated as in Expression (7).

  Thus, since the context model P (X | Y) is a mixture of the prior probability Pr (X) and the candidate model Q (X | Y), the probability is reduced when the amount of learning data is small. A context closer to the context of the target data can be provided while avoiding a context model in which the distribution is extremely biased.

The occurrence probability model selection unit 107 generates the occurrence probability model P (X | Y _in ) when the reference data Y is the reference data Y _in input from the reference data acquisition unit 106 among the context models P (X | Y). Is input to the entropy encoding unit 108.

The reference data acquisition unit 106 inputs the target data S delayed by 4 bits to the occurrence probability model selection unit 107 as reference data Y _in . The entropy encoding unit 108 performs entropy encoding on the value X _in obtained by dividing the target data S by 4 bits using the occurrence probability model P (X | Y _in ) input from the occurrence probability model selection unit 107. The entropy encoding unit 18 outputs the result of entropy encoding as encoded data E. For entropy coding, any code such as an arithmetic code or a Huffman code may be used as long as it uses an occurrence probability. Incidentally, since the delayed target data S 4 bits in the reference data acquisition unit _106, probability model P when entropy coding the _{X in | Y in} the (X _{Y in)} is in the target data S 4 bits immediately before X _in .

  FIG. 3 is a schematic block diagram illustrating the configuration of the decoding device 200. 3, parts corresponding to those in FIG. 2 are denoted by the same reference numerals and description thereof is omitted. The decoding device 200 includes an entropy decoding unit 208 and an occurrence probability generation device 211. The occurrence probability generation device 211 includes an occurrence probability model selection unit 107 and a context model generation device 110.

The entropy decoding unit 208 generates decoded data D by decoding the encoded data E. The entropy decoding unit 208 uses the occurrence probability model P (X | Y _in ) selected by the occurrence probability model selection unit 107 when decoding. In the decoding device 200, the reference data Y _in is 4 bits of the decoded data D decoded immediately before by the entropy decoding unit 208.

As described above, the context model generation device 110 calculates the occurrence ratio of each data value in the learning data and the mixture ratio between the predetermined candidate models and the respective reference data values in the learning data. Determined by referring to the occurrence probability of the data value.
Thereby, a context closer to the context of the target data can be provided.

(Second Embodiment)
In the first embodiment, the encoded data E is output between the encoding device 100 and the decoding device 200, but the encoded data E is output between the encoding device 100a and the decoding device 200a in the present embodiment. In addition, the mixing coefficient a hat is multiplexed and transmitted. FIG. 4 is a schematic block diagram showing the configuration of the encoding device 100a. In the figure, portions corresponding to the respective portions in FIG. 2 are denoted by the same reference numerals and description thereof is omitted. The encoding apparatus 100a has substantially the same configuration as that of the encoding apparatus 100 in FIG. 2 except that it further includes a multiplexing unit 109. The multiplexing unit 109 multiplexes the encoded data E generated by the entropy encoding unit 108 and the mixing coefficient a hat determined by the mixing coefficient determination unit 104 and outputs the multiplexed data.

  FIG. 5 is a schematic block diagram showing the configuration of the decoding device 200a. In FIG. 5, parts corresponding to those in FIG. The decoding device 200a includes a buffer unit 101, an entropy decoding unit 208, a separation unit 209, and an occurrence probability generation device 211a. The occurrence probability generation device 211a includes an occurrence probability model selection unit 107 and a context model generation device 110a. The context model generation device 110a includes a prior probability estimation unit 102 and a context model generation 105.

  Separating section 209 separates the data transmitted from encoding apparatus 100a into encoded data E and mixing coefficient a hat. The separation unit 209 inputs the separated encoded data E to the entropy decoding unit 208. The separation unit 209 inputs the mixing coefficient a hat to the context model generation unit 105.

  In the present embodiment, the mixing coefficient a hat is transmitted to the decoding device 200a. For this reason, the decoding apparatus 200a does not need to obtain the mixing coefficient a from the decoded target data like the decoding apparatus 200. Therefore, before encoding the target data, the context estimation unit 103 of the encoding device 100a determines that the entire target data is the learning data L2 and obtains the mixing coefficient a, and has not been encoded as the learning data L2. You may make it include object data.

Thus, the multiplexing unit 109 multiplexes the encoded data E and the mixing coefficient a hat. The separation unit 209 separates the input data into encoded data E and a mixing coefficient a hat.
Accordingly, the decoding device 200a can decode the encoded data without generating the context of the target data and calculating the mixing coefficient when generating the context model.

In each of the above-described embodiments, the target data S stored in the buffer unit 101 is used as learning data. However, the encoding devices 100 and 100a acquire the learning data separately from the target data S. May be.
In each of the above-described embodiments, the Kronecker delta function is used as the candidate model Q (X | Y). However, the candidate model Q (X | Y) is defined by an LUT or the like, and the context model generation apparatus 110 An LUT may be stored.

  6 and 7 are tables showing examples of LUTs that define the candidate model Q. FIG. In the example shown in FIG. 6, when the X value and the Y value match, the probability is “0.7”, and when they do not match, the probability is “0.1”. In the example shown in FIG. 7, the probability is “0.7” when the value of X is one greater than the value of Y, and “0.1” otherwise. However, when Y is “3”, the probability when X is “0” is “0.7”, and the probability when X is another value is “0.1”.

In each embodiment described above, the context model generation unit 105 linearly combines Pr (X) and Q (X | Y) when generating the context model. b) A plurality of candidate models may be linearly combined as Pr (X) + a × Q ₁ (X | Y) + b × Q ₂ (X | Y). In this case, if an equation in which the partial differentiation of equation (4) with respect to a is 0 and an equation in which the partial differentiation of equation (4) with respect to b is 0 are used as equations solved for a and b. The mixing coefficient determination 104 can determine a and b. Similarly, there may be three or more candidate models for linear combination.

When a plurality of candidate models are used, the relationship between Y _n and X _n may be different depending on the candidate model. For example, in the first candidate models, Y _n is 4 bits immediately before the X _n, the second candidate models, Y _n may be 4 bits before 40-bit X _n. Furthermore, the candidate model may have different bit length of Y _n.

  A plurality of candidate models may be prepared and one of them may be selected and used. For example, the mixing coefficient determination 104 may calculate the value of the formula (4) for a plurality of candidate models and select the candidate model having the smallest value.

  Also, the encoding device 100, the context model generation device 110, the occurrence probability generation device 111 in FIG. 2, the decoding device 200 in FIG. 3, the encoding device 100a in FIG. 4, the decoding device 200a in FIG. 5, the context model generation device 110a, A program for realizing the function of the occurrence probability generation device 211a is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed to implement each device. Also good. Here, the “computer system” includes an OS and hardware such as peripheral devices.

  The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

2 described above, the context model generation device 110, the occurrence probability generation device 111, the decoding device 200 in FIG. 3, the encoding device 100a in FIG. 4, the decoding device 200a in FIG. 5, and the context model generation device. Each functional block of 110a and the occurrence probability generation device 211a may be individually chipped, or a part or all of them may be integrated into a chip. Further, the method of circuit integration is not limited to LSI, and implementation using a dedicated circuit or a general-purpose processor is also possible. Either hybrid or monolithic may be used. Some of the functions may be realized by hardware and some by software.
In addition, when a technology such as an integrated circuit that replaces an LSI appears due to progress in semiconductor technology, an integrated circuit based on the technology can be used.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes design changes and the like within a scope not departing from the gist of the present invention.

DESCRIPTION OF SYMBOLS 10 ... Information source encoding / decoding system 100, 100a ... Encoding apparatus 101 ... Buffer part 102 ... Prior probability estimation part 103 ... Context estimation part 104 ... Mixed coefficient determination part 105 ... Context model generation part 106 ... Reference data acquisition part 107 ... Occurrence probability model selection unit 108 ... entropy coding unit 109 ... multiplexing unit 110, 110a ... context model generation device 111 ... occurrence probability generation device 200, 200a ... decoding device 208 ... entropy decoding unit 209 ... separation unit 211, 211a ... occurrence probability Generator

Claims (7)

A context model generation device that generates a context model indicating an occurrence probability of each data value of target data when a reference data value is given,
A prior probability estimator for calculating a first model composed of the occurrence probabilities of each data value in the first learning data representing the target data;
A context estimator that calculates a second model composed of occurrence probabilities of each data value when a reference data value is given in the second learning data representing the target data;
A candidate model comprising the first model and the occurrence probability of each data value when a reference data value is given, wherein a mixing ratio between the predetermined candidate model and the first model is set to the second model A mixture ratio determining unit that is determined with reference to the model;
A context model generation device comprising: a context model generation unit configured to generate a context model in which the first model and the candidate model are mixed at the mixing ratio. The context model generation unit according to claim 1, wherein the context model generation unit includes generating the context model by linearly combining the first model and the candidate model with the mixture ratio. apparatus.   The context model generation apparatus according to claim 1, wherein the candidate model is a Kronecker delta function. A prior probability estimator for calculating a first model composed of occurrence probabilities of each data value in the first learning data representing the target data;
A context estimator that calculates a second model composed of occurrence probabilities of each data value when a reference data value is given in the second learning data representing the target data;
A candidate model comprising the first model and the occurrence probability of each data value when a reference data value is given, wherein a mixing ratio between the predetermined candidate model and the first model is set to the second model A mixture ratio determining unit that is determined with reference to the model;
A context model generating unit that generates a context model in which the first model and the candidate model are mixed at the mixing ratio;
An encoding apparatus comprising: an encoding unit that generates encoded data by encoding the target data using the context model generated by the context model generation unit.   The encoding apparatus according to claim 4, further comprising a multiplexing unit that multiplexes the encoded data generated by the encoding unit and information indicating the mixing ratio determined by the mixing ratio determination unit. A prior probability estimator for calculating a first model composed of occurrence probabilities of each data value in the first learning data representing the target data;
A context estimator that calculates a second model composed of occurrence probabilities of each data value when a reference data value is given in the second learning data representing the target data;
A candidate model comprising the first model and the occurrence probability of each data value when a reference data value is given, wherein a mixing ratio between the predetermined candidate model and the first model is set to the second model A mixture ratio determining unit that is determined with reference to the model;
A context model generating unit that generates a context model in which the first model and the candidate model are mixed at the mixing ratio;
A decoding unit comprising: a decoding unit that decodes encoded data obtained by encoding the target data using a context model generated by the context model generation unit. A separator that separates input data into encoded data obtained by encoding the target data and information indicating a mixing ratio used for encoding the encoded data;
A prior probability estimator for calculating a first model composed of the occurrence probabilities of each data value in the first learning data representing the target data;
A context model that is a candidate model composed of the occurrence probability of each data value when a reference data value is given and is mixed with a predetermined candidate model at the mixing ratio is generated. A context model generation unit;
A decoding unit comprising: a decoding unit that decodes encoded data obtained by encoding the target data using a context model generated by the context model generation unit.

Images