KR20060122684A - Method for encoding and decoding video signal - Google Patents

Abstract

A method for encoding/decoding a video signal is provided to effectively initialize a context model related to a syntax element at the CABAC(Context-based Adaptive Binary Arithmetic Coding). A binarizer(110) produces a binary code or a bin string related to the symbols of the syntax elements with a predetermined manner. A context modeler(120) selects the context model according to the statistics of the latest coded symbols and adaptively changes the internal state of the selected context model. An arithmetic coder(130) encodes the binarized symbols according to the arithmetic coding method on the basis of the changed internal state. A general coding engine(131) produces each bit on the basis of the probability changed by the context modeler.

Description

Method for encoding and decoding video signal

1 is a schematic representation of the encoding algorithm of AVC,

2 schematically illustrates a configuration of a context-based adaptive binary arithmetic coding (CABAC) encoder according to an embodiment of the present invention.

3 illustrates by way of example a relationship between a quality base layer and an FGS enhanced layer.

4 illustrates an example in which the CABAC context model is initialized to a predetermined value when encoding or decoding for each slice including a slice of an FGS enhanced layer starts.

5 and 6 illustrate embodiments of the present invention for initializing a context model for syntax elements of a current FGS slice using the last state of a previous slice (quality base layer or previous FGS enhanced layer),

7 shows an example in which a slice is divided into several NALs,

8 schematically illustrates a configuration of a CABAC decoder according to an embodiment of the present invention.

<Description of the symbols for the main parts of the drawings>

100: CABAC encoder 110: binarization unit

120: context modeler 130: arithmetic coder

131: Generic Coding Engine 132: Bypass Coding Engine

200: Cabac Decoder 210: Context Modeler

220: Arithmetic Decoder 221: Generic Decoding Engine

222: bypass decoding engine 230: inverse binarization unit

The present invention relates to a method for encoding and decoding a video signal.

The scalable video codec (SVC) method encodes a video signal, and encodes at the highest quality, but decodes a partial sequence of the resultant picture sequence (a sequence of intermittently selected frames in the entire sequence). It is a method to allow a certain level of image expression even if used.

A picture sequence encoded in a scalable manner can be represented to a certain degree of image quality by receiving and processing only a partial sequence. However, when a bit rate is lowered, image quality deterioration is large. In order to solve this, a separate auxiliary picture sequence for a low data rate, for example, a small picture and / or a low picture sequence per frame may be provided.

The auxiliary picture sequence is called a base layer, and the main picture sequence is called an enhanced layer or an enhancement layer. The base layer and the enhanced layer encode the same video signal source, and redundant information (redundancy) exists in the video signals of the two layers. Accordingly, when providing a base layer, an interlayer prediction method may be used to increase coding efficiency.

In addition, an enhanced layer may be used to improve the signal-to-noise ratio (SNR) of the base layer, that is, to improve image quality, which may be used for SNR scalability, fine granular scalability (GFS), or incremental refinement (FGS). Progressive Refinement).

According to the FGS, transform coefficients corresponding to each pixel, for example, discrete cosine transform (DCT) coefficients are encoded by dividing the base layer and the enhanced layer according to the resolution of the bit representation, Since the transmission of the enhanced layer is omitted, the bit rate may be lowered while reducing the quality of the decoded image. That is, the FGS is designed to compensate for the loss occurring during the quantization process, and provides high flexibility in controlling the bit rate in response to a transmission or decoding environment.

For example, if the base layer is generated by quantizing the transform coefficients to a quantization step size (aka QP), for example QP = 32, the FGS enhanced layer 1 inverses the original transform coefficients and the quantized coefficients of the base layer. The difference from the quantized transform coefficient is quantized to QP = 26, which is a quantization step size corresponding to a higher quality than QP32. Similarly, the FGS enhanced layer 2 is generated by quantizing QP = 20 with respect to the difference between the transform coefficient obtained by dequantizing the sum of the original transform coefficient and the quantized coefficients of the base layer and the FGS enhanced layer 1.

SVC is a codec extended from AVC (Advanced Video Codec: also referred to as 'H.264'), and FIG. 1 schematically illustrates an encoding algorithm of AVC. AVC decoding proceeds in the reverse order of the encoding process of FIG. 1, except for the motion prediction process.

The input image signal is an image block having spatial and / or temporal motion prediction and motion compensation of the image, and having residual data based thereon. Is generated. The error data of the generated image block is transformed into coefficients by a 2-D transform method such as a discrete cosine transform (DCT) or a wavelet, and the transform coefficients are quantized in an arbitrary quantization step.

The quantized transform coefficients are scanned in the same order as zigzag or modified zigzag, and then by entropy coding methods such as variable length coding (VLC) or arithmetic coding, It is coded into a compressed bit stream that is easy to transmit or store, along with motion vectors, header information, and additional information.

Entropy coding is a lossless coding method that converts a data stream into a compressed bit stream with a considerably reduced data size to facilitate transmission or storage of elements representing a video image.

Entropy coding includes variable length coding (VLC) mode, which converts symbols of data elements into continuous code words of varying length, and binary arithmetic coding (BAC), which converts consecutive symbols into a single decimal point. ) Mode and the like. In addition, entropy coding may adopt a context-based adaptive (CA) coding method that adaptively codes syntax elements included in a slice layer or a layer below the context. .

These methods combine to form a context-based adaptive variable length coding (CAVLC) mode and a context-based adaptive binary arithmetic coding (CABAC) mode. Although CABAC increases the computational amount compared to CAVLC, it is possible to obtain the optimal bits needed to represent each symbol closer to the theoretical maximum compression rate.

Context-based adaptive binary arithmetic coding (CABAC) generates binary code for the codes of a particular syntax element, selects a context model according to the generated binary code, changes the probability of the selected context model, and based on that Generates a bit sequence by coding.

In this context, the context model may initialize an internal state represented by a probability of occurrence of Most Probable Symbol (MPS) and an MPS value at each start of encoding / decoding of each slice, and calculates actual statistics of the syntax element. In order to reflect, the internal state is constantly updated during encoding / decoding based on the binary symbols of the syntax element.

The reason why the context model is initialized to a predetermined value is not to improve coding efficiency, but to be robust and not sensitive to any condition, that is, robustness. By initializing each context model each time a slice is started, the slice can be decoded independently, regardless of the loss or corruption of the previous slice.

An object of the present invention is to provide a method for efficiently initializing a context model for a syntax element in relation to an FGS enhanced layer and a base layer in CABAC.

In order to achieve the above object, a method of encoding a video signal according to an embodiment of the present invention may include: modeling the element so that a compressed bit stream may be generated by reflecting actual statistics on an element representing the video signal. Wherein the modeling comprises initializing a model for the element when the slice starts, but for an element common to the previous slice (s) associated with the current slice. Initializing based on the last updated model in any previous slice; And updating a model for the element based on the binary symbols of the element.

In the above embodiment, the slice in which the element is initialized based on the model of the previous slice is characterized in that the slice is generated to compensate for the loss occurring in the quantization process for the previous slice.

If there are multiple slices generated to compensate for the loss that occurs during quantization, the model for the elements of the current slice is based on the last updated model in the previous slice or in the first slice, among the multiple slices. Characterized in that it is initialized.

The embodiment further comprises the step of recording information indicating a method for initializing the model for the element, wherein the information comprises: initializing the model for the element to a predetermined value, or immediately before. Initializing based on the last updated model in the slice, or initializing based on the last updated model in the first slice, wherein the information is a slice header, picture header, GOP header, or sequence header It is characterized by being recorded in at least one of the.

According to another embodiment of the present invention, a method of encoding a video signal may include: performing modeling on the element to reflect actual statistics on an element representing the video signal; And generating a bit stream for the element based on the generated model, wherein the performing of the modeling includes: a model for the element when a network abstract layer (NAL) starts; Initializing, but initializing, based on a model last updated in the previous NAL, an element common to the previous NAL associated with the current NAL; And updating a model for the element based on the binary symbols of the element.

In the above embodiment, the model for the element of the current NAL is characterized in that it is initialized based on the last updated model in the immediately preceding NAL partitioned from the same slice.

In addition, the embodiment further comprises the step of recording in the NAL header information indicating whether the model for the element is initialized to a predetermined value or based on the last updated model in the immediately preceding NAL. .

According to an embodiment of the present invention, a method of decoding an encoded bit stream includes: modeling an element representing the video signal based on the encoded bit stream; And generating a continuous symbol for the element based on the generated model, wherein the modeling includes initializing a model for the element when a slice starts. Initializing, based on the last updated model in any previous slice, for an element common in the previous slice (s) associated with the current slice; And updating a model for the element based on the binary symbols of the element.

In the above embodiment, the slice in which the element is initialized based on the model of the previous slice is characterized in that the slice is generated to compensate for the loss occurring in the quantization process for the previous slice.

In addition, the information indicating whether to initialize the model for the element to a predetermined value, based on the last updated model in the previous slice, or based on the last updated model in the first slice includes a slice header, And is read from at least one of a picture header, a GOP header, and a sequence header.

According to another embodiment of the present invention, a method of decoding an encoded bit stream includes: modeling an element representing the video signal based on the encoded bit stream; And generating a continuous symbol for the element based on the generated model, wherein the modeling includes initializing a model for the element when NAL starts. Initializing, based on a model last updated in any previous NAL, an element common to a previous NAL associated with the current NAL; And updating a model for the element based on the binary symbols of the element.

In this embodiment, the model for the element of the current NAL is initialized based on the model last updated in the immediately preceding NAL partitioned from the same slice, and the model for the element lasted to a predetermined value or last in the immediately preceding NAL. The information indicating whether to initialize based on the updated model is read from the NAL header.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

2 schematically illustrates the configuration of a context based adaptive binary arithmetic coding (CABAC) encoder 100. The CABAC encoder, for symbols in the syntax element, in the manner defined for the syntax element, the binarization unit 110 for generating a binary code or a binary string (bin string), for each bin in the generated bin string, A context modeler 120 which selects a context model according to statistics of recently coded symbols and also adaptively changes an internal state (probability of occurrence of MPS or LPS and MPS value) for the selected context model, And an arithmetic coder 130 for encoding the binarized symbol according to an arithmetic coding method based on the changed internal state.

The arithmetic coder 130 generates a regular coding engine 131 for generating each bit based on a probability changed by a context modeler (an occurrence probability of the MPS and an internal state represented by the MPS value), and Internally, there is a bypass coding engine 132 that performs arithmetic coding on syntax elements for which the probability of occurrence of bit values of 0 and 1 is nearly equal so that there is no benefit of context modeling.

The syntax elements of the slice layer or layers below can be coded through CABAC. In addition, syntax elements that can be coded through CABAC include type information (mb_type) related to a macro block, coded_block_pattern (CBP) information indicating a 4x4 block having non-zero data, information related to quantization, reference index information, and a motion vector. Related information, quantized transform coefficient information, and the like.

The syntax element is assigned a binary code (bins) to each symbol in a predetermined manner (or given conversion table). For other syntax elements, they are binarized in different ways or by different conversion tables.

The bins generated from the symbols of the syntax element are applied to the arithmetic coder 130. When the bin values have a probability of being evenly distributed without biasing to either 0 or 1, the arithmetic coding has little benefit, so the detour coding When applied to the engine 32 and the bin value has a probability of biasing to either 0 or 1, it is applied to the context modeler 120 to undergo a modeling process.

The context modeler 120 selects a context model (probability function) for each applied bin based on bin values of the same syntax element of the adjacent macro block and recent bin values of the syntax element of the current macro block, Initialize or update the probability of the MPS value and the LPS (or MPS) in the selected probability function (probability table) and provide it to the arithmetic coder 130.

In the H.264 standard, a context model to be used for each syntax element and a process of updating a probability of LPS according to bin values in each context model are predefined. In addition, each probability function associated with different context models to be used in the H.264 standard is initialized to a predetermined value at the beginning of each slice, and the MPS value and LPS probability based on each applied bin and arithmetic coded bit value. Is updated.

3 illustrates an example of a relationship between a quality base layer and an FGS enhanced layer.

In the example of FIG. 3, the quality base layer is composed of I (intra), B (bi-predicted), and P (predicted) slices, and the FGS enhanced layer is composed of two layers. The encoded bit stream is transmitted in the following order.

I0-FGS00-FGS01-P1-FGS10-FGS11-B2-FGS20-FGS21-P3-FGS30-FGS31-B4-FGS40-FGS41-P5-FGS50-FGS51-B6-FGS60-FGS61-

As shown in FIG. 4, when encoding or decoding for each slice including a slice of the FGS enhanced layer starts, the CABAC context model is initialized to a predetermined state.

Unlike other slices, slices of the FGS enhanced layer are only used to improve SNR and therefore cannot be decoded without the data of the quality base layer. Also, FGS data decoded without a quality base layer is useless. In other words, the decoding of the FGS enhanced layer depends entirely on the quality base layer.

While encoding or decoding the quality base layer or the previous FGS enhanced layer, each context model is updated to reflect the actual statistics of the signal. Since the FGS enhanced layer is an SNR-enhanced layer for the same picture, the signal characteristics of the FGS enhanced layer are very similar and closely related to those of the quality base layer (or the previous FGS enhanced layer).

For example, a syntax element such as mb_qp_delta that indicates the difference between the representative quantization step size of the slice and the quantization step size for that macroblock, or CBP (coded_block_pattern) that indicates whether there is data in the block, may be used for quality base layer and FGS enhancement. In the de-layer, they have similar values.

Thus, rather than initializing the context model for any syntax element of a slice of an FGS enhanced layer (hereinafter simply referred to as an FGS slice) to a predetermined state, the previous slice (quality base layer or previous FGS enhanced layer) It may be more advantageous to initialize the context model for the corresponding syntax element of the current FGS slice using the last state.

In the present invention, when initializing a context model using a previous slice, as shown in FIG. 5, the present invention may be initialized using a last state of a previous FGS slice, or as shown in FIG. 6. Initialize using the last state.

In addition, the present invention allows this initialization method to be adaptively used, for example, using the init_cabac_FGS_idc syntax. The init_cabac_FGS_idc syntax specifies the following initialization method to be used in the context model of the current FGS slice.

i) same method as the quality base layer

ii) initializing using the last state of the previous FGS slice,

iii) Initialize using the last state of the quality base layer.

The init_cabac_FGS_idc syntax may be defined in a slice header. In general, the video sequence from pixel to sequence is hierarchical structure of pixel-> block-> sub macro block-> macro block-> slice-> picture-> sequence. Accordingly, the syntax may include a sequence parameter set (SSP) in a sequence header, a picture parameter set (PPS) in a picture header, or a group of picture (GOP) header including a slice. It may be defined within.

In the CABAC encoder 100 according to the present invention, the context modeler 120, at the beginning of the FGS slice, determines the last state (probability function, and MPS value and the probability function) of the slice of the previous FGS slice or the quality base layer. LPS probability), and initializes the context model for the syntax element.

On the other hand, in addition to the video sequence layer, additional hierarchical structures (layers) such as packetization and channel coding associated with the delivery of auxiliary data streams such as audio and video data are needed. Video compression tools such as motion prediction, 2-D transform, quantization, scan, and entropy coding basically operate on the slice layer or layers below it.

Accordingly, data related to the slice layer or a layer below it is identified as a video coding layer (VCL), and data related to the layer above it is identified as a network abstract layer (NAL). The VCL and the NAL may be transmitted together as part of one bit stream or may be transmitted separately. Specific VCL bit streams may be included in different NALs designed to fit various transport frameworks, such as broadcast, wireless, or media.

For example, as shown in FIG. 7, the quantized transform coefficients are zigzag scanned in macroblock units (or block units).

In this case, the operation of scanning up to the pixel with the nonzero bit in the current macroblock and the pixel with the nonzero bit in the next macroblock is performed in one cycle (cycle #n) to the last macroblock of the slice. Then, for each macro block, from the first macro block to the last macro block, the scan from the last pixel scanned in the previous cycle to the pixel with the next nonzero bit is performed in the next cycle (cycle # (n + 1)). do.

In this case, each cycle may be transmitted through a different NAL. That is, one slice may be divided into several NALs (NAL unit # 0, NAL unit # 1,) and transmitted as shown in FIG. 7. As described above, when one slice is divided into several NALs, contents of a corresponding slice header are attached to each NAL, and parameters, that is, syntax elements, in the slice header attached to each NAL may be different.

Entropy encoding and decoding after scanning is performed in units of NAL, and an initialization operation on syntax elements is also performed every time NAL starts. It is not efficient in terms of arithmetic coding that it is contained in different NALs but initialized to a predetermined value each time a NAL is started for related, ie, syntax elements of the same slice.

Therefore, in the present invention, when initializing a syntax element while starting the NAL, the context model of the syntax element associated with the previous NAL is initialized using the last state updated in the previous NAL.

In the CABAC encoder 100 according to the present invention, when the NAL starts, the context modeler 120 initializes the context model by receiving the last state of the previous NAL with respect to syntax elements related to the previous NAL divided from the same slice. do.

The data stream encoded by the method described so far is transmitted by wire or wirelessly or via a recording medium.

8 schematically illustrates a configuration of the CABAC decoder 200 of FIG. 8. The context modeler 210 performs a context modeling operation on a syntax element based on a coded and input bit stream, and delivers it to the arithmetic decoder 220, and the arithmetic decoder 220 codes based on a state of a context model. And generate a bin string from the input bit value, and the debinarization unit 230 converts the bin string into a symbol of the original syntax element.

When the context modeler 210 performs an initialization operation on a syntax element while starting a slice or NAL, the context modeler 210 may include a slice of a base layer or a previous FGS slice, for a previous slice or a syntax element associated with the NAL. Or use the last state of the previous NAL.

The decoding apparatus including the CABAC decoder of FIG. 8 may be mounted in a mobile communication terminal or the like or in an apparatus for reproducing a recording medium.

As described above, preferred embodiments of the present invention have been disclosed for the purpose of illustration, and those skilled in the art can improve, change, and further various embodiments within the technical spirit and the technical scope of the present invention disclosed in the appended claims. Replacement or addition may be possible.

Therefore, it is possible to efficiently initialize the context model for syntax elements in CABAC and improve coding efficiency.

Claims (18)

And performing modeling on the element to generate a compressed bit stream reflecting actual statistics on an element representing an image signal, Here, the step of performing the modeling, When a slice starts, initializing a model for the element, for an element common to the previous slice (s) associated with the current slice, based on the last updated model in any previous slice; And Updating the model for the element based on the binarized symbols of the element. The method of claim 1, The modeling operation on an element uses a binary symbol of the element to determine the probability function for each binary number of the symbol, the most probable bit value (MPS), and the bit opposite to the MPS or the MPS in the probability function. The method of encoding a video signal, characterized in that the process of selecting a probability value for the value. The method of claim 1, And a slice in which an element is initialized based on a model of a previous slice is a slice generated to compensate for a loss occurring in the quantization process for the previous slice. The method of claim 3, wherein If there are a plurality of slices generated to compensate for the loss occurring in the quantization process, the model for the elements of the current slice is initialized based on the last updated model from the previous slice among the plurality of slices. How to encode a signal. The method of claim 3, wherein If there are a plurality of slices generated to compensate for the loss occurring in the quantization process, an image of an element of the current slice is initialized based on a model last updated in the first slice among the plurality of slices. How to encode a signal. The method of claim 3, wherein And recording information indicating how to initialize the model for the element, Here, the information may be initialized to a model for the element to a predetermined value, based on a model last updated in the immediately previous slice, or based on a model last updated in the first slice. A method of encoding a video signal, characterized in that it refers to either. The method of claim 6, Wherein the information is recorded in at least one of a slice header, a picture header, a GOP header, or a sequence header. Performing modeling on the element to reflect actual statistics on the element representing the image signal; And Based on the generated model, generating a bit stream for the element, Here, the step of performing the modeling, Initializing a model for the element when a network abstract layer (NAL) starts, and initializing the element common to the previous NAL associated with the current NAL based on the model last updated in the previous NAL; And Updating the model for the element based on the binarized symbols of the element. The method of claim 8, And the model for the element of the current NAL is initialized based on the last updated model in the immediately preceding NAL partitioned from the same slice. The method of claim 9, And recording in the NAL header information indicating whether the model for the element is initialized to a predetermined value or based on the last updated model in the immediately preceding NAL. Modeling an element representing the video signal based on the encoded bit stream; And Generating a continuous symbol for the element based on the generated model, Here, the step of performing the modeling, When a slice starts, initializing a model for the element, for an element common to the previous slice (s) associated with the current slice, based on the last updated model in any previous slice; And Updating the model for the element based on the binarized symbols of the element. The method of claim 11, And the slice in which the element is initialized based on the model of the previous slice is a slice generated to compensate for the loss incurred in the quantization process for the previous slice. The method of claim 12, If there are a plurality of slices generated to compensate for the loss occurring in the quantization process, the model for the elements of the current slice is initialized based on the last updated model from the previous slice among the plurality of slices. The decoded bit stream. The method of claim 12, Encoding of the elements of the current slice is initialized based on the last updated model in the first slice among the plurality of slices, if there are a plurality of slices generated to compensate for the loss occurring in the quantization process The decoded bit stream. The method of claim 12, Slice header, picture header, indicating whether to initialize the model for the element to a predetermined value, based on the last updated model in the immediately previous slice, or based on the last updated model in the first slice. And reading from at least one of a GOP header or a sequence header. Modeling an element representing the video signal based on the encoded bit stream; And Generating a continuous symbol for the element based on the generated model, Here, the step of performing the modeling, When a NAL starts, initializing a model for the element, but for an element common to a previous NAL associated with a current NAL based on a model last updated in any previous NAL; And Updating the model for the element based on the binarized symbols of the element. The method of claim 16, And the model for the elements of the current NAL is initialized based on the last updated model in the immediately preceding NAL partitioned from the same slice. The method of claim 16, And decoding information from the NAL header indicating whether the model for the element is initialized to a predetermined value or based on the last updated model in the immediately preceding NAL. How to

Images