KR102609632B1 - Methdo and apparatus for processing a video signal - Google Patents

Abstract

Provides an intra prediction method and device through reference sample smoothing.

Description

Video signal processing method and apparatus {METHDO AND APPARATUS FOR PROCESSING A VIDEO SIGNAL}

The present invention relates to a video signal processing method and device.

Video images are compressed and encoded by removing spatial-temporal redundancy and inter-view redundancy, and can be transmitted through communication lines or stored in a format suitable for storage media.

The present invention seeks to improve the coding efficiency of video signals.

In order to solve the above problems, the present invention provides an intra prediction method and device through reference sample smoothing.

The video signal processing method and device according to the present invention can improve video signal coding efficiency through reference sample smoothing.

In video encoding, a predicted image is generated using intra prediction of the current picture or inter prediction. The predicted image is encoded as a calm residual image from the original image. The closer the image quality of the predicted image is to the original, the smaller the sample value of the residual image is, increasing the encoding efficiency.

The input video signal is divided into several blocks, and among them, the blocks that share the prediction mode (Intra Mode, Inter Mode, or Skip Mode) are determined and divided into coding units ( It is called a Coding Unit (CU). The size of the coding unit may be a square block between 64x64 and 8x8, a square block with a size of 128x128, 256x256 or more, or a rectangular block of any size. You can optionally use any number of CU sizes among square blocks ranging in size from 256x256 to 8x8. For example, in some sequences, only CUs of size 256x256, CUs of 128x128, and CUs of 32x32 may be used, and the size of the CU to be used in the current picture or sequence may be signaled in the sequence header or picture header.

The picture is divided into square or non-square basic blocks and the coding process is performed, which are called Coding Tree Units (CTU). In the sequence parameter set (SPS), it can be signaled whether it is a square CTU or a non-square CTU, and the CTU size can be signaled. In the case of a non-square CTU, both the CTU width and the CTU height may be signaled. Alternatively, the size or shape of the forward and/or non-square CTU may be derived based on the number of vertical/horizontal boundaries dividing the picture.

The CTU can be divided into a quad tree/binary tree structure and encoded by forming a coding unit, or divided into a hierarchical structure and encoded.

The coding unit is encoded using any one of skip mode, intra mode, and inter prediction types. Once the coding unit is determined, the prediction unit (PU) is determined. The size of the prediction unit is equal to or smaller than the size of the coding unit. The prediction unit is determined through the prediction splitting (PU splitting) process of the coding unit. And prediction division of the coding unit is performed according to the partition mode (Part_mode) of the coding unit. That is, the size or shape of the prediction unit is determined depending on the partition mode of the coding unit. In the case of inter-screen prediction (Inter Mode), it is divided into a total of 8 prediction blocks, as shown in Figure 1.

Figure 1 Used for cross-screen prediction partition type

For intra-screen prediction (Intra mode), PART_2Nx2N and PART_NxN are used. For partition type PART_NxN, it is used only when the size of the CU is the same as the size of the smallest CU defined in the slice. Generally, the size of the PU is defined as 64x64 to 4x4, and in the case of inter mode prediction, the size is 4x4 to reduce memory bandwidth when performing motion compensation (MC). PU is not used.

One. Intra How to generate prediction samples

In the case of intra-screen prediction, the boundary sample of a neighboring block that has already been encoded is used as a reference pixel for intra-screen prediction. There are 35 modes in which HEVC (High Efficient Video Codec) version 1 generates a predicted image within the screen using reference pixels. The 35 modes consist of DC mode, planar mode, and 33 directional prediction modes. Figure 2 shows 33 directional prediction modes.

Figure 2 Inside the screen Prediction mode

The intra-screen prediction mode of the chroma component is determined as shown in Table 1 below based on the intra-screen prediction mode used in the luma component. That is, DC mode, planar mode, horizontal prediction mode, vertical prediction mode, diagonal prediction mode, and a mode that reuses the mode used in the luma component can be used. .

Alternatively, more directional prediction modes may be used than the 33 directional prediction modes defined in HEVC version 1. That is, an extended directional prediction mode may be defined by further subdividing the angle of the directional prediction mode, and directional prediction with a predetermined angle may be performed using at least one of a predefined number of directional prediction modes. The mode can also be derived and used from the decoding device.

For example, Figure 3 shows an extended intra prediction mode, and the total number of intra prediction modes may be 67. It can also be configured in DC mode, planner mode, and 65 directional prediction modes.

Figure 3 65 directions Intra prediction mode

The extended intra prediction mode can also be used for both luma and chroma components. Different numbers of intra prediction modes may be used for the luma component and the chroma component. For example, the luma component may use 67 intra prediction modes, and the chroma component may use 35 intra prediction modes. Alternatively, intra prediction may be performed using different numbers of intra prediction modes depending on the chroma format. For example, in the case of 4:2:0 format, intra prediction can be performed using 67 intra prediction modes in luma, and 35 intra prediction modes can be used in chroma, and in the case of 4:4:4 format, luma and Intra prediction can also be used using 67 intra prediction modes on all chroma components.

Alternatively, intra prediction may be performed using different numbers of intra prediction modes depending on the size of the block (size or shape of PU or CU). That is, intra prediction may be performed using 35 intra prediction modes or 67 intra prediction modes depending on the size or shape of the PU or CU. For example, if the size of the CU or PU is smaller than 64x64 or is an asymmetric partition (partition mode such as PART_nLx2N in inter prediction mode), intra prediction can be performed using 35 intra prediction modes, and the CU or PU If the size is equal to or larger than 64x64, intra prediction can be performed using 67 intra prediction modes.

For another example, Intra_2Nx2N may allow a 65-way intra prediction mode, and Intra_NxN may only allow a 35-way intra prediction mode.

As described above, the number of extended intra prediction modes can be selectively determined by considering at least one of the component type (Luma component or chroma component), chroma format, block size, and partition mode. However, it is not limited to this, and the number of intra prediction mode candidates used for intra prediction of the decoding target block among intra prediction modes also considers at least one of component type (Luma component or chroma component), chroma format, block size, and partition mode. This can be determined selectively.

For each sequence or slice unit, the size of the block applying a different number of intra prediction modes can be applied differently. For example, in one slice, intra prediction is performed using 67 intra prediction modes on a PU or CU larger than 64x64, and in another slice, intra prediction is performed using 67 intra prediction modes on a PU or CU larger than 32x32. It can also be done. Information indicating this, log2_extended_intra_mode_size_minus4, can be signaled in a sequence parameter set or slice header, etc.

If log2_extended_intra_mode_size_minus4 is 0, it indicates that intra prediction can be performed using 67 intra prediction modes on a CU or PU larger than 16x16, and if log2_extended_intra_mode_size_minus4 is 1, 67 intra prediction modes can be performed on a CU or PU larger than 32x32. Indicates that intra prediction can be performed using .

When performing intra prediction using an extended intra prediction mode, determine a plurality of intra prediction mode candidates (mpm) among the extended intra prediction modes, and selectively use one of the plurality of intra prediction mode candidates (e.g. For example, the intra prediction mode of the current block (selected based on the signaled mpm_idx) can be derived. At this time, the number of the plurality of intra prediction modes is 3, 4, 5, 6, or more most probable modes can be used fixedly or adaptively.

If the extended intra prediction mode and the 35 intra prediction modes defined in HEVC version1 are optionally used, convert the intra mode of the surrounding block to the index of the extended prediction mode or the index of the 35 intra prediction modes defined in HEVC version1 After converting to , the most probable mode can be derived. A pre-defined table may be used for index conversion, or a scaling operation based on a predetermined value may be used. Here, the pre-defined table may define a mapping relationship between different intra prediction mode groups (for example, the extended intra prediction mode and the intra prediction mode defined in HEVC version 1).

For example, if the left prediction unit uses 35 intra prediction modes such as HEVC version 1, and the intra prediction mode index is 10 (horizontal mode), the mode index is set to 16, which is the horizontal mode index of the extended intra prediction mode. After conversion, it can be used to derive the most probable mode, etc.

For another example, if the upper prediction unit uses an extended intra prediction mode and the intra prediction mode index is 50 (vertical mode), after converting the mode index to 26, which is the vertical mode index of HEVC version1, the most probable It can also be used when deriving a mode, etc.

In the Luma component, intra-screen prediction is performed after filtering the reference pixel used for intra-screen prediction. Whether or not filtering is applied is determined depending on the intra-screen prediction mode and the size of the transformation unit (TU). In DC mode or when the TU size is 4x4, filtering is not performed. In other cases, filtering is performed when the difference between the intra-screen prediction mode and the vertical prediction (or horizontal prediction) mode is greater than a pre-defined threshold. The threshold values are shown in Table 2 below.

8x8 transform 16x16 transform 32x32 transform Threshold 7 One 0

For filtering, a bilinear interpolation filter, a 3-tap filter, or a filter with more taps can be optionally used. Bilinear interpolation filter can use predetermined weights (e.g., [1/2, 1/2], [1/3, 2/3], [1/4, 3/4]), and is a 3-tap filter. A filter with filter coefficients of [1,2,1] can be used.

When using intra-screen prediction, the predicted image is created using boundary samples of surrounding blocks, so the image quality of the predicted image is lowered compared to inter-screen prediction.

After performing the intra prediction mode to generate a prediction sample, there may be a case where the difference between the sample values of the surrounding block and the current block becomes larger than the difference between the samples of the actual block. To compensate for this, filtering is performed on samples of neighboring blocks and/or samples on the boundary of the current block to generate a final prediction signal, thereby reducing the difference in sample values between blocks and increasing coding efficiency.

For example, in vertical mode, the final predicted sample can be obtained using the difference between the left neighboring sample p(-1,y), which is closest to the current sample, and the upper left neighboring sample p(-1,-1), using the following equation ( The prediction signal can be obtained as shown in 1).

P'(0,y) = P(0,y) + (( p(-1,y) -p(-1,-1)) >> 1 for y=0…N-1 (1)

In horizontal mode, the final predicted sample can be obtained using the difference between the upper neighboring sample p(x,-1), which is closest to the current sample, and the upper left neighboring sample p(-1,-1), as shown in Equation (2) below: A prediction signal can be obtained.

P'(x,0) = p (x,0) + ((p(x,-1) -p(-1,-1))>>1 for x= 0 .. N-1 (2)

Figure 4 Intra prediction in mode Examples of current and surrounding samples

In the case of the current vertical prediction mode or horizontal prediction mode, the final prediction sample is generated using the difference information of neighboring blocks only on one line at the block boundary, and multiple lines (e.g., multiple columns, The final prediction sample may be generated using difference information of neighboring blocks (a plurality of rows).

Specifically, for example, in vertical mode, the final predicted sample can be obtained using the difference between the left neighboring sample p(-1,y), which is closest to the current sample, and the upper left neighboring sample p(-1,-1). At this time, the difference may be added to the predicted value of the current sample, or the difference may be scaled to a predetermined value and then added to the predicted value of the current sample. A predetermined value used for scaling may be determined differently depending on the row and/or column. As an example, a prediction signal can be obtained as in the following equations (3) and (4).

P'(0,y) = P(0,y) + (( p(-1,y) -p(-1,-1)) >> 1 for y=0… N-1 (3)

P'(1,y) = P(1,y) + (( p(-1,y) -p(-1,-1)) >>2 for y=0… N-1 (4)

In horizontal mode, the final predicted sample can be obtained using the difference between the upper neighboring sample p(x,-1), which is closest to the current sample, and the upper left neighboring sample p(-1,-1), which is as described above in vertical mode. same. As an example, the prediction signal can be obtained as in the following equations (5) and (6).

P'(x,0) = p (x,0) + ((p(x,-1) -p(-1,-1))>>1 for x= 0 .. N-1 (5)

P'(x,1) = p (x,1) + ((p(x,-1) -p(-1,-1))>>2 for x= 0 .. N-1 (6)

Alternatively, if the prediction mode index of the directional prediction mode is 2 or 34, the current prediction sample obtained after performing intra prediction based on the surrounding samples and the directional intra prediction mode as shown in Figure 5 below is divided into the current prediction sample and at least one located in the lower left corner. The final predicted sample can be obtained through filtering based on the predicted/restored sample. Here, the lower left prediction/reconstruction sample may belong to the line preceding the line to which the current prediction sample belongs, may belong to the same block as the current sample, or may belong to a neighboring block adjacent to the current block.

Sample filtering can be performed only on the line located at the block boundary, or it can be performed on multiple lines as shown in the figure. Different filter coefficients can be used for each line. For example, for the first line on the left closest to the block boundary, you can use the (1/2,1/2) filter, for the second line, you can use the (12/16, 4/16) filter, and for the third line, you can use the (1/2,1/2) filter. In this case, the (14/16, 2/16) filter is used, and in the case of the fourth line, the (15/16, 1/16) filter can also be used.

Figure 5 Prediction sample filtering method

Alternatively, if the prediction mode index during directional prediction mode is a value between 3 and 6 or between 30 and 33, filtering can be performed only at the block boundary as shown in Figure 6, and prediction samples can be generated using 3 tap filtering. The final prediction sample can be generated by filtering the lower left sample of the current sample, the lower left sample of the lower left sample, and the current sample using a 3-tap filter. The positions of neighboring samples or reference samples used for filtering may be applied differently for each prediction mode.

Depending on the intra prediction mode, the filter coefficients of the 3-tap filter may be applied differently.

Figure 6 Prediction sample filtering method

Different filtering can be applied depending on whether the surrounding block is in inter mode or intra mode. For example, if the surrounding block is encoded in intra mode, a filtering method that gives more weight to the current pixel can be used than if it is encoded in inter mode. For example, if the intra prediction mode is 34 and the neighboring blocks are encoded in inter prediction mode on the first line at the block boundary, the (1/2,1/2) filter is used, and the neighboring blocks are encoded in intra prediction mode. If encoded, it can also be encoded using a (4/16, 12/16) filter.

The number of lines performing prediction sample filtering may be applied differently depending on the size of the prediction unit or coding unit. For example, if the PU size is less than or equal to 32x32 as shown in the left image of Figure 7, filtering is performed on only one line at the block boundary, and if the PU size is larger than 32x32 as shown in the right image of Figure 7, filtering is performed on the block boundary. Filtering can also be performed on multiple lines.

Figure 7 Based on PU size filtering area

2. Intra block copy

An image of a block (hereinafter referred to as a reference block) similar to the current block in the already restored image in the screen may be used as a prediction image. This method is called intra block copy (IBC). In intra block copy, in order to specify the position of the reference block, the difference in position of the reference block from the current block can be expressed as a vector like a motion vector, and this is called a block vector. Figure 8 shows the intra block copy method. A prediction block vector may be generated through prediction encoding, and only the difference vector between the block vector and the prediction block vector may be encoded, or the block vector of the current block may be encoded without prediction encoding. Here, the prediction block vector may be derived from the block vector of a neighboring block adjacent to the current block. Alternatively, it may be derived from a block vector in an LCU containing the current block, or may be derived from a block vector in an LCU row containing the current block.

In particular, when the image contains many characters such as the Korean alphabet or the alphabet, if the previous block contains characters to be encoded in the current block, intra block copy can be used to improve encoding performance.

Figure 8 Intra block copy

3. Weighted prediction method using surrounding samples

Even though the corresponding blocks of the current block and the previous frame are similar, changes in the brightness of the previous frame and the current frame occur, so that they are not encoded with intra prediction or inter prediction, or the image quality of the predicted image encoded with intra prediction or inter prediction is relative. Low cases may occur. If a brightness change occurs throughout the frame, a weighted image P' can be obtained by adding a weight w and an offset f to the image P encoded/decoded by intra-prediction or inter-prediction as shown in equation (7) below. Here, the weight w and offset f may be applied to the prediction sample generated through intra-prediction or inter-prediction, or may be applied to the final reconstructed sample.

P'=w x P + f (7)

At least one of weight w or offset f may be signaled in the sequence header or slice header.

An entire slice or sequence may experience changes in lighting compared to the previous slice or sequence, or only some areas of a slice or sequence may experience changes in lighting compared to the previous slice or sequence. In this case, at least one of the weight w or the offset f may be signaled for each block (e.g., a block such as a coding unit or prediction unit) in the encoding step. At least one of the weight w or the offset f may be signaled regardless of the coding mode of the block, or may be selectively signaled in consideration of the coding mode of the block. For example, if the coding mode of the block is inter mode, the weight w or offset f may be signaled, and otherwise, it may not be signaled. Here, the inter mode may include at least one of skip mode, merge mode, or intra block copy mode. However, intra block copy is not limited to being included in inter mode and may be classified as intra mode.

Weights can also be derived using lighting changes between a specific type of first template adjacent to the current block and a second template adjacent to the corresponding previous block. The second template may include an unavailable sample. In this case, the available sample can be copied and used in the position of the unavailable sample, or the available sample can be derived and used through interpolation between a plurality of available samples. The available sample used at this time may belong to the second template or a neighboring block. At least one of the coefficients, shape, or number of taps of the filter used for interpolation may be variably determined based on the size and/or shape of the template.

For example, if the neighboring sample of the current block is y _i (i is 0 to N-1) and the neighboring sample of the reference block is x _i (i is 0 to N-1), the weight w and offset f are as follows. It can be derived as follows.

Using a specific type of template adjacent to the current block, the weight w and offset f can be derived by obtaining the minimum value of E(w,f) in the following equation (8).

(8)

The method of finding the minimum value of equation (8) can be modified as equation (9).

(9)

From equation (9), equation (10) deriving the weight w and equation (11) deriving the offset f can be obtained.

(10)

(11)

4. template How to configure

Figure 9 is an example of configuring a template consisting of neighboring samples of the current block.

Figure 9 template Day Example

As shown in the left picture of Figure 9, all neighboring samples adjacent to the current CU can be configured as a template, or, as shown in the central picture, the template can be configured with some samples sub-sampled from all neighboring samples adjacent to the current CU. The central figure is an example of 1/2 sub-sampling, and the template can be constructed only from the gray portion of the sample. Instead of 1/2 sub-sampling, the template can be configured using 1/4 sub-sampling or 1/8 sub-sampling. As shown on the right side of Figure 9, a template can be constructed by excluding the sample located in the upper left corner from all neighboring samples adjacent to the current CU. Although not shown in Figure 8, considering the location of the current CU in the picture or coding tree block (Largest Coding Unit), a template consisting of only samples located on the left can be used, or a template consisting of only samples located at the top can be used.

Alternatively, the template can be constructed by expanding the number of adjacent neighboring samples, as shown in Figure 10. Although the template in Figure 9 uses only neighboring samples adjacent to the boundary of a prediction unit or coding unit, it can also be composed of a neighboring sample adjacent to the coding unit boundary and adjacent neighboring samples, as shown in one embodiment of Figure 10.

Figure 10 template Day Example

As shown in the left image of Figure 10, all neighboring samples belonging to two adjacent lines from the boundary of the prediction unit or coding unit can be used as a template, or a template can be constructed by sub-sampling from the template in the left image as shown in the middle image. As shown in the right picture of Figure 10, a template can be constructed by excluding the four samples in the upper left corner from the template shown in the left picture. Although not shown in Figure 10, considering the location of the current CU in the picture or coding tree block (Largest Coding Unit), a template consisting of only samples located on the left may be used, or a template consisting of only samples located at the top may be used.

Different templates can be configured depending on the size or shape of the prediction unit or coding unit (square, non-square, or asymmetric prediction unit partition).

For example, as shown in Figure 11, the sub-sampling rate of the template can be applied differently depending on the size of the coding unit. For example, in CUs whose size is less than or equal to 64x64, a 1/2 sub-sampling template is configured as shown in the left image of Figure 11, and in CUs whose size is larger than or equal to 128x128, a 1/4 sub-sampled template is constructed as shown in the right image of Figure 11. -You can configure a sampling template.

Figure 11 Based on CU size template How to choose

As another embodiment, a template that expands the number of adjacent neighboring samples according to the size of the CU can be used, as shown in Figure 12.

Figure 12 Based on CU size template How to choose

A plurality of template candidates available in a sequence or slice can be determined, and any one of them can be selectively used. The plurality of template candidates may be configured as templates of different shapes and/or sizes. The shape, form, and/or size of the template may be signaled in the sequence header or slice header. In the encoder or decoder, an index value can be assigned to each template candidate. Additionally, the syntax type_weight_pred_template_idx can be encoded to identify the template candidate used in the current sequence, picture, or slice among a plurality of template candidates. The decoder can selectively use the template candidate based on the syntax type_weight_pred_template_idx.

For example, as shown in Figure 13, the template shown in the middle picture of Figure 9 is assigned to 0, the template shown in the right picture of Figure 9 is assigned to 1, the template shown in the middle picture of Figure 10 is assigned to 2, and , the template shown on the right side of Figure 10 can be assigned to 3, and among them, the template used in the sequence can be signaled.

Figure 13 template

5. non-square in the block template How to configure

If the coding unit is in the form of a non-square block or weighted prediction is performed using a prediction unit that allows non-square blocks, the sub-sampling rate on the long and short sides so that the total number of templates can be 2^N. You can also configure the template by applying it differently. For example, as shown in Figure 14, a template can be constructed by performing 1/2 sub-sampling on the short side and 1/4 sub-sampling on the long side.

Figure 14 Non-square block template Day Example

6. Sub-block unit Intra How to generate a prediction signal

Even if an intra prediction signal is generated using the directional intra prediction method, since the prediction signal is generated using only the surrounding pixels of the current block, there are cases where the prediction image does not reflect the characteristics of the original image, and in this case, the residual signal (register As the value of (Dual) becomes relatively large, the disadvantage of increasing the amount of bits during encoding may occur.

For example, when an edge exists within a prediction unit or an object that was not present appears around the boundary of the current block, there is a part where the difference between the directional intra prediction image and the original image becomes relatively larger depending on the position within the current block. . In this case, coding efficiency deteriorates because residuals with many high-frequency components are generated.

In particular, the residual signal in an area relatively far from the boundary of the block contains a lot of high-frequency components. In order to improve the image quality of the prediction signal in an area relatively distant from the block boundary, a method of updating the prediction signal on a sub-block basis may be used.

For example, a second prediction signal can be obtained by applying a predetermined offset to each sub-block to the first prediction signal generated by performing directional intra prediction or intra block copy. If the current block (eg, coding block) is divided into multiple sub-blocks and encoded, a predetermined offset may be selectively applied to each sub-block. To this end, whether a predetermined offset is applied to each subblock may be signaled, or whether the offset may be applied may be determined depending on the location of the corresponding subblock within the current block. As shown in Figure 15, in subblocks close to the block boundary, the first prediction signal is used as the final prediction signal, and in subblocks far from the block boundary, a predetermined offset is applied to the first prediction signal to generate the second prediction signal, which is It can be used as the final prediction signal. At this time, the second prediction signal may be derived by adding or subtracting a predetermined offset from the first prediction signal.

Figure 15 Sub-block unit Intra Predictive signal generation way to work Example

For another example, a second prediction signal may be generated by applying different offsets to the first prediction signal on a sub-block basis, as shown in Figure 16.

Figure 16 Sub-block unit Intra Predictive signal generation pattern work Example

The syntax is_sub_block_refinement_flag, which indicates whether to apply the sub-block unit intra prediction signal generation method for each prediction unit or coding unit, can be signaled. If the value of is_sub_block_refinement_flag is 1, the sub-block unit intra prediction signal generation method can be applied.

When using the sub-block unit intra prediction signal generation method, it is possible to signal which pattern of sub-block intra prediction signal generation method to use. The decoder may generate a second prediction signal by applying an offset to the first prediction signal of at least one sub-block on a sub-block basis according to a predefined pattern. Specifically, for example, when using index 0, as shown in the first picture of Figure 15, the upper sub-block uses the first prediction signal as is, and the lower sub-block applies a predetermined offset f to the first prediction signal to perform second prediction. A signal can be generated. Alternatively, as shown in the first picture of Figure 16, the upper sub-block can generate the second prediction signal by applying a predetermined offset h to the first prediction signal, and the lower sub-block can generate the second prediction signal by applying a predetermined offset f to the first prediction signal. 2 A prediction signal can be generated.

Alternatively, a specific pattern may be selectively used to generate a sub-block unit prediction signal based on the intra prediction mode used to generate the first prediction signal. For example, if it is a horizontal intra prediction mode or an intra prediction mode with a direction similar to the horizontal direction intra prediction mode (e.g., the intra prediction mode index is 6 to 14), index2 or index3 in Figure 16 can be used, In the case of an intra prediction mode with a directionality similar to the vertical intra prediction mode or the horizontal intra prediction mode (for example, when the intra prediction mode index is 6 to 14), index0 or index1 in Figure 16 can be used.

As shown in Figure 17 below, subblocks within an intra prediction unit or coding unit may be composed of different sizes, or an intra block may be composed of three subblocks to generate a second prediction signal.

After grouping the sub-block unit intra prediction patterns in Figures 15 to 17 into several categories, the category identification information and the index indicating the pattern belonging to the category can be signaled. For example, index 0 to index 3 in Figure 17 may be set to category 1, index 4 to index 7 in Figure 17 may be set to category 2, and index 8 to index 12 may be set to category 3. After signaling which category to use among these, the index of the pattern used within the intra prediction unit or coding unit may be signaled. Alternatively, only the index that specifies the pattern may be signaled without separate category grouping, or the pattern may be derived from the partition mode of the prediction unit or coding unit. Whether to signal an index that specifies a pre-defined pattern or derive it from a partition mode can be determined at the picture, slice, or block level, and a separate flag for this may be signaled.

Figure 17 Subblock unit prediction signal pattern work Example

Figures 15 to 17 are just one example of a sub-block unit intra prediction pattern, and may be determined differently within the range obvious to those skilled in the art.

The number of intra prediction patterns in subblock units can be determined on a sequence or slice basis, and the number can be signaled in the sequence header or slice header.

Based on the size of the coding unit or the size of the prediction unit, a different number of sub-block-level intra prediction patterns or a different number of sub-blocks can be applied.

For example, if the size of the CU is 64x64 or larger, the pattern in Figure 18 is used to generate a sub-block unit intra prediction signal, and if the CU size is smaller than 64x64, the pattern in Figure 15 to Figure 17 is used to generate the sub-block unit intra prediction signal. A unit intra prediction signal can also be generated.

Figure 18 Sub-block unit prediction signal pattern work Example

As another example, as shown in Figure 19, an intra prediction signal in sub-block units may be generated using patterns with different sizes and/or shapes between sub-blocks. A sub-block unit intra prediction signal can also be generated using at least one of the patterns shown in Figure 19 below.

Figure 19 Sub-block unit prediction signal pattern work Example

The offset h or f or g or i used to generate the second prediction signal on a sub-block basis may be signaled on a slice basis or a CU basis, or may be derived from neighboring samples around the current block. For example, the offset for each subblock can be derived by considering the distance between the current sample at a certain location (left, top border, top left corner sample, etc.) within the current block. For example, the offset may be determined in proportion to a value representing the distance between a predetermined position in the current block and the current sample.

7. In video coding of symbol context How to initialize probability index

A bitstream can be generated by binarizing symbols such as transform coefficients, motion vector difference, and intra-slice syntax and performing arithmetic coding. Each symbol selects a probability index for each symbol based on the context determined by considering the value of the same symbol in the surrounding block, information on the surrounding block, or location within the current block, and the cumulative statistics of the internal symbol Compression performance can be improved by recalculating the probability according to the value of the encoded symbol and using arithmetic coding. The arithmetic coding method used in HEVC is called CABAC (Context Adaptive Binary arithmetic Coding).

To encode a symbol, the following process can be performed.

(1) Binarization

If the target symbol to be encoded is not a binary symbol (a symbol whose symbol values consist of only 0 and 1) (for example, transform coefficient and motion vector difference), the target symbol is converted to a binary symbol. In HEVC, symbols can be binarized using unary and truncated unary binarization methods. When a symbol is binarized, the bit with 0 or 1 among the mapped codewords is called a bin.

Table 3 below shows the unary binarization method.

Symbol binarization 0 0 One 10 2 110 3 1110 …

Table 4 below shows the Truncated unary binarization with cMax = 6 method.

Symbol binarization 0 0 One 10 2 110 3 1110 4 11110 5 111110 6 111111

(2) Select context model

Context model refers to the probability model for each symbol. The probability of 0 or 1 occurring in a bin may be different for each context model. In HEVC, there are approximately 400 independent contexts for various symbols.

When a slice starts, the context probability index pStateIdx for each context of each symbol can be initialized based on the quantization parameter Qp value or slice type (I slice, P slice, or B slice).

When using Tile, the context probability index pStateIdx for each context of each symbol at the front of each tile can be initialized based on the quantization parameter Qp value or slice type (I slice, P slice, or B slice).

(3) Arithmetic coding

Each symbol performs an arithmetic encoding process on a context model basis. Even if the same symbol is used in a different context, probability update and bitstream encoding do not affect each other. If the probability of each symbol is high, fewer bits are used, and if the probability of each symbol is low, lower bits are used. For example, if the probability value is high, a symbol with 10 bins may be encoded with fewer than 10 bits.

The arithmetic coding method divides the symbol to be encoded into sub-intervals between [0, 1) based on the probability value of the symbol, then selects the number that can be expressed with the smallest bit among the real numbers belonging to the sub-interval, and encodes the coefficient. It's a method. When dividing the section between [0.1) into sub-intervals, if the symbol probability is large, a long sub-interval can be assigned, and if the symbol probability is low, a small sub-interval can be assigned.

For example, assuming that the probability of occurrence of 1 is 0.2 and the probability of occurrence of 0 is 0.8, as shown in Figure 20, the arithmetic coding procedure for symbol 010 is as follows.

a) Encode the first 0 in Symbol 010. Since the probability of occurrence of 0 is 0.8, the interval is updated from [0,1) to [0-0.8).

b) Encoding the second 1 in Symbol 010. Since the probability of occurrence of 1 is 0.2, the interval is updated from [0-0.8) to [0.64-0.8)

c) Encoding the third 0 in Symbol 010. Since the probability of occurrence of 0 is 0.8, the interval is updated from [0.64-0.8) to [0.64-0.768)

d) Since the smallest real number representation in [0.64-0.768) is 0.75 = (1/2)x1 +(1/2)^2 ×1, the binary expression 11 excluding 0 is encoded.

Figure 20 Arithmetic coding example

(4) Probability update of symbols

Based on the context, MPS (Most Probable Symbol, the most frequently occurring symbol between 0 and 1) and LPS (Least Probable Symbol, the least frequently occurring symbol among the two binary numbers 0 and 1) of each bin are set to their initial values based on the Qp value. This is set. Depending on whether the currently encoded bin is MPS or LPS, the MPS occurrence probability and LPS occurrence probability values of that symbol may change.

For example, if the binarization value of the bin of the symbol to be currently encoded is equal to MPS, the MPS probability value of this symbol becomes larger and the LPS probability value becomes smaller, and if it is equal to LPS, the MPS probability value of this symbol becomes smaller and the LPS probability value becomes smaller. can grow. In CABAC, a total of 64 MPS occurrence probabilities and a total of 64 LPS occurrence probabilities can be configured. It can be defined as an index pStateIdx that represents the probability of occurrence of a symbol, and as the value of pStateIdx increases, the probability of MPS occurrence increases.

As shown in Table 5, if the MPS symbol is encoded in the context probability index pStateIdx, which represents the probability value of the current context, pStateIdx can be updated to the index corresponding to transIdxMPS. For example, when the pStateIdx value is 16, if the MPS symbol is encoded, pStateIdx can be updated to index 17 indicated by transIdxMPS, and if the LPS symbol is encoded, pStateIdx can be updated to index 13 indicated by transIdxLPS. When the pStatIdx value is 0 and the LPS symbol is encoded, the MPS and LPS symbols are exchanged. (When pStateIdx is 0, the probability of MPS occurrence is set to 0.5. When the LPS symbol is encoded, the frequency of LPS becomes higher than that of MPS, so the MPS and LPS symbol values are changed.)

7.1 plural context based on probability index context How to initialize a model

In HEVC, it is initialized and used as the initial context probability index value of each context on a slice or tile basis. If a symbol is encoded with an initial value already set at the beginning of a slice, encoding performance may deteriorate because it does not properly reflect the probability of the actual symbol. However, because the context probability index is initialized on a slice basis, there is also the advantage of being able to decode the current slice even if the previous slice or frame has not been transmitted.

While decoding the previous slice, the context probability index accumulated at a certain point in time (for example, when encoding a block located in the middle of the scan order) can be used as the initial value of the context probability index of the current slice. The context probability value or context probability index value accumulated in the previous slice may be directly encoded in the slice header, etc.

Alternatively, at least one context can be initialized using multiple context probability indices. For example, a random context ctx may have a different context probability index initial value for each slice. Which probability index to select among a plurality of context probability indexes for each slice can be transmitted in the slice header, and in the decoding step, decoding can be performed based on the context probability index information transmitted in the slice header. Regardless of the slice type, multiple context initial values (InitValue) or context probability indices (pSateIdx) can be used.

A context initial value (InitValue) is set to derive a different probability index for each symbol's context. InitValue can be used to derive variables m and n, where variables m and n can be defined as values used to derive the variable preCtxState representing the previous context state. Based on the derived variable preCtxState, the MPS value and the initial context probability index value pStateIdx can be derived. For example, the context probability index initial value pStateIdx of the corresponding context can be derived as shown in the following equations (12), (13), and (14).

slopeIdx = initValue >> 4

offsetIdx = initValue & 15 (12)

m = slopeIdx * 5 - 45

n=(offsetIdx << 3)-16 (13)

preCtxState=Clip3(1,126,((m*Clip3(0,51,SliceQp _Y )) >> 4)+n)

valMps=(preCtxState <= 63)? 0:1

pStateIdx = valMps ?( preCtxState - 64 ) : ( 63 - preCtxState ) (14)

InitValue can be set differently for each slice type, and the syntax num_cabac_init_idx_minus1, which indicates how many context initialization values are present in the slice or sequence header or picture header, can be signaled. For example, based on the num_cabac_init_minus1 value, Table 6 to Table 6 At least one of 7 may be optionally used.

The index cabac_init_idx that specifies the context initial value (InitValue) used in the slice may be signaled in the slice header. The context initial value (InitValue) corresponding to cabac_init_idx can be determined based on a table that defines the mapping relationship of at least two of cabac_init_idx, ctxIdx, or initValue.

For example, as shown in Tables 6 and 7, the contexts of the syntax cbf_luma, which indicates whether a non-zero coefficient exists in the transform block of the luma component, have different initial values, that is, different context probability index pStateIdx, if the cabac_init_idx value is different. It may be possible.

cabac_init_idx 0 One 2 3 4 ctxIdx 0 One 0 One 0 One 0 One 0 One initValue 111 141 153 111 153 111 168 224 95 79

cabac_init_idx 0 One 2 3 4 5 ctxIdx 0 One 0 One 0 One 0 One 0 One 0 One initValue 111 141 153 111 153 111 168 224 95 79 63 31

Multiple context models may have multiple context probability indices (i.e., multiple pStateIdx initial values) even when the slice Qp is the same. For example, when the slice Qp value of a slice is sliceQpY, the context probability index pStateIdx can be obtained from the initValue determined for each context. At this time, pStateIdx can be recalculated by applying a specific offset to pSateIdx according to the cabac_init_idx value, or per context. pSateIdx can also be derived by applying a specific offset to the determined initValue.

Rather than applying it to all symbols, multiple context probability indices may be used only for specific symbols, such as transform coefficient, motion vector difference, or reference index. Specifically, for example, a specific symbol such as transform coefficient, motion vector difference, or reference index may have multiple pStateIdx initial values.

Alternatively, different context probability indices may be used depending on the location of the spatial region of the block to be encoded. For example, as shown in Figure 21, different context probability indices pStateIdx can be set depending on the scan order of the slice. At this time, the pStateIdx value set may be selected as the value most similar to the probability index prevPstateIdx of the collocated region of the previous slice.

Figure 21 Multiple areas by area context How to choose a model

The spatial area in which the context probability index is to be initialized is called the context initialization area. The context initialization area may be rectangular in shape or of any size. Information indicating the context initialization area or context initialization area can be signaled in the slice header, etc.

For example, when the context initialization area has a rectangular shape, the unit for initializing the context probability index may be indicated based on the syntax num_row_ctu_minus1, which indicates the number of CTU (Coding Tree Unit) columns included in the context initialization area. Specifically, in Figure 21, num_row_ctu_minus1 can be set to 1, and the area containing the actual two rows of CTUs can be set as the context initialization area.

7.2 tile unit context How to initialize a model

In HEVC, a slice is the basic unit that can independently perform entropy decoding. A slice can be divided into multiple slice segments, and the slice shape does not necessarily have to be rectangular. In the case of a slice segment, it may be composed of multiple Coding Tree Units (CTU).

In the case of a tile, it is a square-shaped area with multiple CTUs, and entropy decoding can be performed on a tile-by-tile basis, and has the advantage of being able to parallelize decoding multiple tiles simultaneously during implementation.

As shown in Figure 22, a tile can have multiple slice segments, and one slice segment can exist within the same tile. Alternatively, there may be multiple tiles within a slice.

Independent slice segments and dependent slice segment(s) are combined to form one slice, and there can be multiple slices within a tile, as shown in Figure 23. Alternatively, there may be multiple tiles within a slice.

painting 22 tiles sliced segment example

Figure 23 Example of a tile with more than 2 slices

When using Tile, the context model is initialized on a tile basis. Depending on the position of the tile, a different context initial value (InitValue) or a different context probability index pStateIdx can be used. That is, the probability index pStateIdx of the same context may have different values depending on the tile. The index tile_cabac_init_idx, which specifies the context initial value (InitValue) used for each tile, may be signaled in the slice segment header, etc. The context probability index pStateIdx can be derived from the initValue specified from tile_cabac_init_idx used in each tile.

The context-specific probability index pStateIdx of each tile may be derived based on the context probability index pStateIdx or the InitValue of the corresponding context of the corresponding tile (collocated tile) of the previous frame, or may be derived by selecting from a plurality of InitValues or a plurality of pSateIdx defined for each context. You can use it. When selecting and using a plurality of pStateIdx defined for each context, the pSateIdx value or context initial value (InitValue) used for each tile may be signaled.

Figure 24 In tile symbol How to initialize probability index

8. In video coding affine motion model( Apine motion model) using Inter Predictive coding method

Representative motion models include translation, rotation, and affine motion models. The translation motion model can represent linear movement as shown in Figure 25-a, and the movement of the object can be expressed as 2D coordinates or motion vectors (MvX, MvY). A motion model in which the movement of an object can be expressed in the form of angle transformation, as shown in Figure 25-b, is called a rotation motion model. As shown in Figure 25-c, when the movement of an object changes non-linearly and has different translation and rotation depending on the part of the object, it is called an affine motion model. Affine motion model can express movement, rotation, size adjustment, etc., and can be expressed using six parameters as shown in equation (14) below.

Figure 25 motion Model example

x'=ax + by + e

y'=cx+dy+f (14)

Here, x represents a pixel belonging to the current block, x' represents a pixel belonging to a reference block (corresponding block of the reference picture), and the difference between the pixel of the current block and the pixel of the reference block is expressed as (V _x , V _y ).

V _x = x - x'

V _y = y - y' (15)

Using equations (14) and (15), the affine motion vector (V _x , V _y ) can also be expressed as follows.

V _x = (1-a)x-by-e

V _y = (1-d)y-cx-f (16)

8.1 Simplified affine motion model encoding method

While the affine motion model can express complex motion models, it has the disadvantage of lowering encoding efficiency as the number of parameters to be encoded increases compared to the translation motion model or rotation motion model.

As shown in the left picture of Figure 26, an affine motion model can be used using the motion vectors (v0, v1, v2) at the upper left, upper right, and lower left corners of the block. In this case, the affine motion vector is as shown in Equation (17) below. can be created.

To overcome this, an affine motion model can be modeled using a motion vector corresponding to the corner position of the block. Here, the corner location of the block may include at least one of the upper left corner, upper right corner, lower left corner, or lower right corner. For example, as shown in Figure 26, an affine motion model can be modeled using motion vectors at the upper left corner and upper right corner of the block.

As shown in Figure 26, it can be expressed as a transformed rectangular block with the pixel pointed to by the upper left motion vector and the pixel pointed to by the upper right motion vector as one line segment.

(17)

Figure 26 Simplified affine motion model

Instead of using the three upper-left, upper-right, and lower-left corners of the block, an affine motion model can be approximated by using the motion vectors from the upper-left and upper-right corners of the block.

At this time, v _2x = v _0x + v _0y - v _1y , v _2y = -v _0x + v _0y + v _1x Equation (17) can be approximated as equation (18) using . The affine motion vector (V _x , V _y ) can be expressed as follows.

(18)

Depending on the size or shape (square/non-square) of the block (prediction unit or coding unit), the position of the corner that induces the affine motion model can be set differently. For example, if the prediction unit is in the Nx2N form as shown in Figure 27, an affine motion vector can be derived using the upper left and lower left corners of the prediction unit. The affine motion vector can also be derived using equation (18).

Figure 27 Non-jeongbang In a prediction unit of the form affine Example of motion model derivation method

The motion vector of the upper left corner of the block is called the first affine vector, and the motion vector of the lower left or upper right corner of the block is called the second affine vector.

8.2 Apine Movement Compensation ( Affine motion compensation)

As shown in Figure 28, after dividing the current block into several sub-blocks, an affine motion vector is generated at a specific position (at least one of the center position, upper left position, or lower left position) of each sub-block. An inter prediction signal can be generated by performing sub-block unit motion compensation. Here, the affine motion vector may be obtained using equation (17).

Figure 28 sub-block unit affine motion field

The number of sub-blocks within a block may be set differently depending on the size or shape of the prediction unit or coding unit. For example, if the prediction unit size is 16x16, affine motion compensation can be performed using four 8x8 unit sub-blocks, and if the prediction unit size is 64x64 or 32x32, a 16x16 unit sub-block can be used. Thus, affine motion compensation can be performed.

Alternatively, affine motion compensation can be used using a sub-block with a predetermined size regardless of the size of the prediction unit, and a syntax indicating the size of the sub-block can be signaled in the slice header.

8.3 Apine Inter mode ( affine inter mode)

The case where the prediction unit/coding unit performs affine motion compensation using an affine motion vector is defined as an affine inter mode. Affine inter mode can be selectively applied depending on the size or shape of the coding unit/prediction unit. For example, affine inter mode may be performed only when the coding unit/prediction unit is square and its size is 16x16 or more.

The first affine vector and the second affine vector may be predicted and used from neighboring blocks of the current block. For example, as shown on the left of Figure 29, the first affine vector can be predicted using at least one of A, B, and C located in the surrounding blocks (the first affine motion vector predictor), and the second affine vector can be predicted using at least one of A, B, and C located in the neighboring blocks. The in vector can be predicted using at least one of D and E (second affine motion vector predictor) located in neighboring blocks.

A first affine difference vector that is a difference value between a first affine vector and the first affine motion vector predictor may be encoded, and similarly, a first affine difference vector that is a difference value between a second affine vector and the second affine motion vector predictor may be encoded. 2 Affine difference vectors can be encoded.

For another example, the first affine vector can be predicted using at least one of B, C, and F located in neighboring blocks, and the second affine vector can be predicted using at least one of A, D, and E located in neighboring blocks. You can also make predictions using one.

Figure 29 Apine motion Vector prediction example

When the values of the first affine vector and the second affine vector are the same, or when prediction is performed from the same block, the second affine vector is predicted from a vector at a different location, or a predetermined affine vector is added to the first affine vector. The second affine vector may be derived by adding or subtracting the offset or applying a predetermined scaling factor. For example, as shown in the left picture of Figure 29, the first affine vector derives an affine vector from A, B, and C, and the second affine vector derives an affine vector from D and E, the value of which is In cases where they are the same, an affine vector can be derived from F. Flag or index information indicating whether the affine vectors are identical may be signaled, and information specifying a corner position with the same affine vector may be signaled. Alternatively, the second affine vector may be encoded based on differential coding based on the first affine vector. That is, the second affine vector may not be encoded as is, and only the difference vector with the first affine vector may be encoded. In this case, the decoding device may restore the second affine vector using the first affine vector and the difference vector. For convenience of explanation, only the first affine vector and the second affine vector are mentioned, but it is not limited thereto. That is, the third affine vector and the fourth affine vector may be used, and the above-described encoding/decoding method may be applied in the same way.

Different affine prediction candidate lists can be constructed depending on the shape of the prediction unit (e.g., square, symmetric, etc.). For example, in the case of a non-square shape as shown in the right picture of Figure 29, the first affine vector can be derived from C, B, and F located in the surrounding blocks, and the second affine vector can be derived from C, B, and F located in the surrounding blocks. Affine vectors can be derived from A, D, and E.

9. Intra in prediction reference Sample smoothing method

As shown in Figure 30, in order to perform intra prediction, reference samples P(-1,-1), P(-1,y) (0<= y <= 2N-2), P(x,-) located at the boundary of the prediction block are used. 1) (0 <= x <= 2N-2) is used. As described above in Chapter 1, filtering can be selectively performed depending on the intra prediction mode or TU size, and this is called reference sample smoothing.

painting 30 intra for prediction reference sample example

When smoothing the reference sample by applying a (1,2,1) filter, the final reference sample to be used for intra prediction can be derived using the following equations (19) to (21).

P(-1,-1) = ( P(-1,0) + 2P(-1,-1) + P(0,-1) + 2) >> 2(19)

P(-1,y) = ( P(-1,y+1) + 2P(-1, y) + P(-1,y-1) + 2) >> 2 (20)

P(x,-1) = ( P(x+1,-1) + 2P(x,-1) + P(x-1,-1) + 2) >> 2 (21)

In equation (20), y is an integer between 0 and 2N-2, and in equation (21), x is an integer between 0 and 2N-2.

Alternatively, reference sample smoothing may be performed on a coding unit basis by selectively using at least one of a plurality of filters. Specifically, for example, at least one of the (1,2,1) 3-tap filter or (2,3,6,3,2) 5-tap filter may be selectively used within the coding unit.

The following equations (22) to (24) represent reference sample smoothing using the (2,3,6,3,2) filter.

P(-1,-1)=( 2P(-2,0)+3P(-1,0)+6P(-1,-1)+3P(0,-1)+2P(0,-2) +8) >>4 (22)

P(-1,y)=(2P(-1,y+2)+3P(-1,y+1)+6P(-1,y)+3P(-1,y-1)+2P(- 1,y-2)+8)>>4(23)

P(x,-1)=(2P(x+2,-1)+3P(x+1,-1)+6P(x,-1)+3P(x-1,-1)+2P(x -2,-1)+8)>>4(24)

In equation (23), y is an integer between 0 and 2N-2, and in equation (24), x is an integer between 0 and 2N-2.

Alternatively, reference sample smoothing may be performed selectively using different smoothing filters depending on the location of the reference sample. For example, the (1,2,1) filter may be applied to reference samples at the border of an intra block, and (2,3,6,3,2) may be applied to other reference samples.

For example, reference samples P(-1,-1) , P(-1,0) , P(-1,1) , … , P(-1,N-1) and P(0,-1), P(1,-1), … , P(N-1,-1) performs reference sample smoothing as shown in equations (19) to (21) using the (1,2,1) filter, and other reference samples use equations (23) to (24) ), reference sample smoothing can be performed.

The filter used for reference sample smoothing is not limited to the (1,2,1) or (2,3,6,3,2) filters described above, and other types of smoothing filters may be used.

Alternatively, reference picture smoothing may be performed selectively using different filters based on the transform type used in the intra block.

For example, as shown in Figure 31, if the intra block is coded using DCT, reference sample smoothing is performed using the (1,2,1) filter, and if it is coded using DST, (2,3,6) ,3,2) Reference sample smoothing can also be performed using a filter.

Figure 31 reference Based on sample location reference Sample smoothing method

As another example, if the intra block is coded using DCT or DST, then reference sample smoothing is performed using the (1,2,1) filter, and if the intra block is coded using the KLT transform, then ( 2,3,6,3,2) Reference sample smoothing can also be performed using filters.

Alternatively, reference picture smoothing may be selectively performed using different filters in consideration of at least one or both of the transform type used in the intra block and the location of the reference sample.

For example, if the intra block is encoded using DCT, the reference samples P(-1,-1) , P(-1,0) , P(-1,1) , … , P(-1,N-1) and P(0,-1), P(1,-1), … , P(N-1,-1) performs reference sample smoothing as shown in equations (19) to (21) using the (1,2,1) filter, and other reference samples use equations (23) to (24) ), reference sample smoothing can be performed. When encoded using DST, the reference samples P(-1,-1), P(-1,0), P(-1,1), … , P(-1,N-1) and P(0,-1), P(1,-1), … , P(N-1,-1) can perform reference sample smoothing as shown in equations (22) to (24) using the (2,3,6,3,2) filter, and other reference samples are used in equations Reference sample smoothing can also be performed as in (20) to (21).

Alternatively, reference picture smoothing can be selectively performed using different filters by comparing the transform type of the block containing the reference sample with the transform type of the current intra block. For example, if the current block and surrounding blocks both use the same transform type, reference sample smoothing is performed using a (1,2,1) filter, and if the current block and neighboring blocks use different transform types, reference sample smoothing is performed. In this case (the current block uses DCT and the block containing the reference sample uses DST), reference sample smoothing can also be performed using the (2,3,6,3,2) filter.

Figure 32 Transform type of surrounding blocks

Alternatively, reference sample smoothing can be selectively performed using different filters based on the transform type of the surrounding block. In other words, reference sample smoothing can be performed using different filters depending on the transform type of the block to which the reference sample belongs. It may be possible.

For example, as shown in Figure 31, the block adjacent to the current intra block and the left/bottom left is a block coded using DCT, and the block adjacent to the top/top right is a block coded using DST. Reference sample smoothing is performed using a (1,2,1) filter on the reference sample adjacent to and a (2,3,6,3,2) filter is performed on the reference sample adjacent to the upper/upper right corner. You may.

Reference sample smoothing may be selectively applied based on Coding Tree Unit (CTU) units. For example, reference sample smoothing can be performed using different filters on a CTU basis, as shown in Figure 33.

Figure 33 CTU based on unit reference Sample smoothing example

In this case, the type of filter used for each CTU (the index of the filter used by the corresponding CTU among available filters) can be signaled in the SPS or picture parameter set (PPS), and the corresponding pictures all use the same filter. It is also possible to signal whether to use it or not.

Claims (12)

determining an intra-screen prediction mode for the current block;
Deriving reference samples of the current block;
Obtaining prediction samples for the current block using the reference samples;
determining whether to perform correction on the prediction samples; and
If it is decided to perform the correction, obtaining a corrected prediction sample,
When the intra-screen prediction mode is a vertical directional prediction mode, the corrected prediction sample is obtained based on a weighted value of the difference between the left reference sample and the upper left reference sample located on the same horizontal line as the prediction sample. become,
An image decoding method, wherein the weight for the difference value is set differently depending on the column to which the prediction sample belongs. delete According to claim 1,
When the prediction sample belongs to the second leftmost column of the current block, the weight for the difference value is 1/2 compared to the weight for the difference value when the prediction sample belongs to the second leftmost column of the current block. Video decoding method. delete delete According to claim 1,
When the intra-prediction mode is a vertical directional prediction mode, the correction is performed on a plurality of columns in the current block. determining an intra-screen prediction mode for the current block;
Deriving reference samples of the current block;
Obtaining prediction samples for the current block using the reference samples;
determining whether to perform correction on the prediction samples; and
If it is decided to perform the correction, obtaining a corrected prediction sample,
When the intra-screen prediction mode is a vertical directional prediction mode, the corrected prediction sample is obtained based on a weighted value of the difference between the left reference sample and the upper left reference sample located on the same horizontal line as the prediction sample. become,
Characterized in that the weight for the difference value is set differently depending on the column to which the prediction sample belongs.
Video encoding method. delete According to clause 7,
When the prediction sample belongs to the second leftmost column of the current block, the weight for the difference value is 1/2 compared to the weight for the difference value when the prediction sample belongs to the second leftmost column of the current block. video encoding method. delete delete According to clause 7,
When the intra-prediction mode is a vertical directional prediction mode, the correction is performed on a plurality of columns in the current block.