WO2016048186A1

WO2016048186A1 - Rapid selection of spatial prediction mode in hevc coding system

Info

Publication number: WO2016048186A1
Application number: PCT/RU2014/000848
Authority: WO
Inventors: Максим Павлович ШАРАБАЙКО; Олег Геннадьевич ПОНОМАРЕВ
Original assignee: Закрытое акционерное общество "Элекард наноДевайсез"
Priority date: 2014-09-24
Filing date: 2014-11-10
Publication date: 2016-03-31
Also published as: EA201700050A1; RU2562414C1

Abstract

The invention relates to coding and decoding digital video data. In order to build a list of candidate modes, a module of high-frequency horizontal and vertical components (details) of a Haar stationary wavelet transform of an image of a coded area is analyzed. The process of building a list of candidate modes consists of four stages. The first and second stages involve deciding whether to include a Planar mode and a DC mode in the list. The third stage involves selecting angular candidate modes according to values in an array of vertical details, and the fourth stage involves selecting candidate modes based on an analysis of values of an array of horizontal details H(x,y). The technical result consists in accelerating the coding process.

Description

QUICK SELECTION OF SPATIAL PREDICTION MODE IN HEVC CODING SYSTEM

Technical field

The invention relates to the encoding and decoding of digital video data.

State of the art

HEVC video compression algorithms are based on a few simple ideas. If we take a certain part of the image, then with a high probability near this section in this frame or in neighboring frames there will be a section containing a similar image that differs little in terms of pixel intensity. Thus, to transmit information about the image in the current section, it is sufficient to transmit only its difference from a previously encoded similar section. The process of finding similar areas among previously encoded images is called prediction (from the English prediction). A set of difference values that determine the difference between the current section and the found prediction is called the remainder (from the English residual). Two main types of prediction can be distinguished. In the first of them, Prediction values are a set of linear combinations of pixels adjacent to the current image area on the left and top. Such a prediction is called spatial (from the English Intra Prediction). In the second, linear combinations of pixels of similar sections of images of previously encoded frames are used as a prediction (these frames are called reference frames from the English Reference). Such a prediction is called temporary (from the English Inter Prediction). To restore the image of the current section, encoded with temporal prediction, when decoding, information is required not only about the Residual, but also about the frame number on which a similar section is located, and the coordinates of this section.

At the next coding stage, the Residual values obtained during the prediction are subjected to the two-dimensional cosine Fourier transform with subsequent quantization. Then, the obtained set of quantized spectral coefficients, accompanied by the information necessary to make predictions during decoding, is subjected to entropy encoding.

The basic structural unit in HEVC is the coding unit (CU - abbr. From English coding unit). Inside each such block, areas are selected - prediction blocks (PU - abbr. From English prediction unit). The video frame is divided into CU adaptively, so it is possible to adjust the borders of the CU to the boundaries of the objects in the image, and the embedded CUs form a quad tree.

Within each CU, areas for prediction calculation — the Prediction Unit (PU) —are selected. In spatial prediction, the CU region may coincide with the PU (2Nx2N mode) or may be divided into 4 square PUs half the size (xN mode). The standard defines the minimum and maximum possible sizes of PU - 4x4 and 32x32, respectively.

Spatial prediction is performed in HEVC in one of 35 ways. In this case, the values of pixels - "neighbors" are used, adjacent to the border of the encoded PU on the left and top. Spatial prediction methods in HEVC can be divided into two unequal groups. The first group includes two methods - Planar and DC. In Planar mode, the values used as predictions lie on a plane whose slope is vertical and horizontal directions is determined by the pixels - "neighbors". This mode is designed to predict areas of the PU with a linear change in pixel values in any direction. In DC mode, a single value equal to the arithmetic average of the neighboring pixels is used as a prediction of all pixels from the PU. The purpose of this mode is self-explanatory.

The modes of the second group are called angular. When calculating the value used as a prediction, in all 33 angular modes, the neighboring pixels are shifted in a given (one of 33) directions. If the position of the predicted pixel falls between shifted copies of the "neighbor" pixels, linear interpolation is used to calculate the prediction. The accuracy of estimating the position of the predicted pixel between the shifted "neighbor" pixels is set to 1/32 of the pixel interval.

The adaptability of splitting the maximum possible CU together with a large number of possible prediction modes gives rise to a huge number of possible prediction options for each such CU so that the process of encoding video frames becomes extremely computationally intensive. Reducing the computational cost of coding is possible, first of all, due to the preliminary selection of prediction modes, building a list of candidate modes. After constructing such a list, the prediction mode for each CU is selected only from candidate modes. It is clear that the shorter the list, the more computationally efficient the coding system. On the other hand, such a list should with high probability contain the best prediction mode for each CU, that is, ensuring the minimum difference between the prediction and the encoded pixels and the minimum number of bits representing the CU in the encoded video stream. From the prior art there is a known method of selecting one of the 35 prediction modes during encoding implemented in the encoder (HEVC Test Model NM v. 11.0 [Electronic resource]. - Access mode: https://hevc.hhi.fraunhofer.de/svn/svn_HEVCSoftware / tags / HM-ll. О /) in which iterates over all possible prediction options for each encoded block. In this case, a complete encoding-decoding cycle of each block is performed, which allows the so-called Rate-Distortion Optimization (RDO). In the RDO process, from all possible modes, one is selected that provides the greatest degree of compression of the video data of the encoded block (the lowest rate) at the lowest level of distortion introduced into this data during the encoding (lowest level of Distortion).

The disadvantage of this solution is that this approach is extremely computationally expensive, but provides a guaranteed choice of the best prediction mode.

There is also a method called fast (Zhao, L .; Zhang, L .; Ma, S .; Zhao, D. Fast mode decision algorithm for intra prediction in HEVC. Visual Communications and Image Processing (VCIP). IEEE, 201 1 , pp. 1-4), which is implemented in two stages. At the first stage, the encoded block is predicted by all possible methods and a shortened list of candidate modes is formed. The selection of the best mode from the generated list is carried out in the second stage of the RDO process. The selection of candidate modes at the stage of list formation can be carried out according to various criteria determined by the settings of the coding system. As such a criterion, the minimum value of the sum of the absolute differences between the pixels of the prediction and the original (SAD or Sum of Absolute Differences) may be the minimum value of the sum of the squared differences of the prediction and original pixels (SSE or Sum of Sqared Errors) or the minimum value of the sum of the modules of the Hadamard transform coefficients of the differences of the prediction and original (SATD or Sum of Absolute Transformed Differences).

The disadvantage of this method is that to build a list of candidate modes, it is necessary to predict the encoded area in all 35 ways. The fast method is again based on a complete enumeration of all 35 prediction modes for each coded area. Some acceleration here is achieved not by reducing the number of predicted prediction modes, but by simplifying the procedure for assessing the quality of prediction in each of the modes. Based on such simplified quality assessment procedures, a shortened list of candidate modes is constructed for which a complete RDO-assessment of the quality of the prediction is already carried out, completely similar to that used in the first method.

SUMMARY OF THE INVENTION

The technical result, which is solved using the proposed solution, is to reduce the amount of computation when searching for the optimal coding mode of the block, which allows to speed up the coding process as a whole, due to the preliminary selection of prediction modes and the construction of a list of candidate modes.

The technical result is achieved in that in a method for quickly selecting a spatial prediction mode in the HEVC coding system, which consists in constructing a shortened list of spatial prediction candidate modes with a length of not more than six positions based on the analysis of the high-frequency horizontal and vertical components of the stationary wavelet transform of the Haar encoded image region, according to the proposed solution:

- an array of values of the high-frequency horizontal component of the Haar transform is obtained by subtracting from the value of each pixel of the encoded region the values of the neighboring left pixel;

- the array of values of the high-frequency vertical component of the Haar transform is obtained by subtracting from the value of each pixel of the encoded region the values of the neighboring neighboring top pixel;

- Planar mode is included in the list of candidate modes if all absolute values of horizontal and vertical high-frequency components are less than the quantization step during encoding;

- DC mode is included in the list of candidate modes if the standard deviation of the horizontal and vertical high-frequency components is less than the quantization step,

- two angular prediction modes corresponding to the mutual displacement of the points of intersection of the line of the minimum values of the gradient modulus passing through the maximum point of the array of vertical high-frequency components, with vertical boundaries of the array of high-frequency components are included in the list of candidate modes,

- two angular prediction modes corresponding to the mutual displacement of the points of intersection of the line of the minimum values of the gradient module passing through the maximum point of the array of horizontal high-frequency components, with horizontal boundaries of the array of high-frequency components are included in the list of candidate modes, - the final choice of the prediction mode for the encoded image area from the constructed list of candidate modes is carried out on the basis of standard RDO (rate-distortion optimization) estimates.

The implementation of the method

To build a list of candidate modes, an analysis of the module of high-frequency horizontal and vertical components (details) of the stationary Haar wavelet transform (CVH) of the image of the encoded area is carried out.

To obtain horizontal and vertical details of the SVPK, arrays hC (x, y), x = - \, 0, ..., nT- \, y = l, 0, ..., nT-l and vC (x, y ), ^x - ^~ 1.A ... ^, nG - 1 _} = 1 ... iTA. The values of the elements of these arrays at x = 0, ..., nT-1, y = 0, ..., nT- 1 are equal to the pixel intensities of the encoded PU. The values ^Η ^ ( ^~ ^> Y) 'Y ^{= ■■ •• n} T - ι are _{equal to} the intensities of the pixel “neighbors” adjacent to the PU region on the left, and hC (x - ΐ x = 0 pT - 1

V ')' '"" are equal to the values of the pixel intensity -

"Neighbors" adjacent to the PU on top. Similarly, the values of vC (x, - 1), x = - 1, ..., ^η Τ - 1 are equal to the intensities of the neighboring pixels ”adjacent to the PU region above, and vC (–l, y), y = 0 , ..., nT - i - to the values of the intensity of the neighboring pixels ”adjacent to the PU on the left. For PT, the size of the encoded PU is indicated. Horizontal and vertical details of H and V are formed as:

H (x, y) = \ hC (x, y) - hC (x - 1, y) \, x = 0, nT - 1, y = 0, ..., nT.

V (x, y) = \ vC {x, y) - vC {x, y - l \, x = Q, ... nT, y = 0, ..., ηΤ - ι.

The process of building a list of candidate modes consists of four stages. At the first stage, a decision is made to include Planar mode in the list. This mode is included in the list if: max [maxHix.yXmaxV ix.yn <qStep

\ h.u h.u,

where: qStep is the quantization step,

In the second step, the DC mode is checked. This mode is included in the list if the condition is met:

Σ _{χ ν} Η (χ, y) ^ _xv V (xy) At the third stage, the candidate angular modes are selected by the values in the array of vertical details ^ (x> Y). For this, the position of Ushakh o, ^x max n of the maximum element of the array V (x, y) is determined.

The current position in the _c is set equal to about Usha, HSG - equal ^ ma D. Iteratively for each next column with the number x _{cr =} jc _cr +1, the new value of y _{c is} set equal to the number of the maximum element in this column from the range [Usg-2, Usg ⁺ ^ - _> ] · The position update continues until x _cr <nT and 0 <y _cr <nT-1. The points x _cg , y _cg during iterations pass along the line of the minimum gradient of the values of V x ^, y ^{) to the} right of the maximum element of this array. A similar iterative process allows us to trace the line of the smallest gradient of values (Ck.u ^{) to the} left of the position * max. This process begins by setting the current position of Ya equal to Ushakh o, x _c \ - equal to goax n. Iteratively for each next column with the number x _c i = x _c i -1, the new value y _c i is set equal to the number of the maximum element in this a column from the range [Us1 ^{~ 2} -Us- ^{+ 2} 1. Updating the position of x _with [, y _c i continues until x _with [> 0 and 0 <y _cr <nT-1.

If both positions y _c and y _{c \} fall on the same block boundary, the results of such a search are incorrect. In this case, the search is performed up and down in accordance with the algorithm of the fourth step applied to the details of V (x, y).

The relative position of the two points ^x a, Us1 and * cg, Usg determine the direction in which the IntraPredAngle angular prediction mode parameter and the prediction mode number are determined. The values of the variables are determined by the following expressions:

(UUss11 - UU h if 1u _s1 - Ust! S fccr - Xrf l

71HP = \

^X * cr - Xcb ^if tycl - Usr '> 1 cr - * ε -

1, eoinliy _cZ - y _cr \> c ,. - x _cl \ _^ 3

dir = _n .,. . . taphi = mn If dir ^{= 0} , then intraPredAnglel = α _ϊ? JntraPredAngle2 = a _{i + i}} where £ -32, -26, -21, -17, -13, -9, -5, -2,0,2,5,9,13,17,21,26,32) _and ^a c -5 tgphi <a _{i + 1} a, the} numbers of the corresponding prediction modes are in the range 2 - 17.

If dir = 1 ₅ then IntraPredAngiel = - “IntraPredAngle2 = -α _{ί + 1 5} where α ε £ -32, -26, -21, -17, -13, -9, -5, -2,0,2, 5,9,13,17,21,26,32) and q ^ tgphi <α _{ί + ί? and the} numbers of the corresponding prediction modes are in the range 18 - 34.

The fourth stage is similar to the third, but the candidate modes are selected on the basis of the analysis of the values of the array of horizontal details H (> y). To do this, determine the position of Ushah o, ^x shzkh the maximum element of the array H ⁽ xy ⁾ . The current position of Ust is set equal to Utah o, * ^' st - equal to * pgah. Iteratively for each next line with the number Ust = Ust + ¹ , the new value% cg is set equal to the number of the maximum element in this line from the range i ^x cr - 2.% cg +. The update of the position continues until Set pT and 0 <* cr <pT - 1. During the iterations, the points% cg, cvr go along the line of the minimum gradient of the values of H ⁽ x ^, y ⁾ below the position of the maximum element of this array. A similar iterative process allows us to trace the line of the smallest gradient of the values of H x.yU over the position ^x max max, utax. This process begins by setting the current position of U a equal to Ushakh, ^x ci - equal to * max. Iteratively for each next line with the number Y a = Y a - ¹ , the new value x _with 1 is set equal to the number of the maximum element in this line from the range [ _c i - 2> Xci ^~ l ^{~ 2} ]. The updating of the position y _c i continues while Y a ^ o _and O <x _c1 <nT - i

If both positions x _cg and x _c fall on the same block boundary, the results of such a search are incorrect. In this case, the search is performed left and right in accordance with the algorithm of the third stage, applied to the array of values of horizontal details Ж * -). The relative position of the found two points ^x c1, Wa and Hsg, Ust specify the direction in which the parameter of the angular prediction modes IntraPredAngle and the number of the prediction mode are determined. We define the values of the variables by the following expressions;

₌ (V a - Ust if! Y _c1 - y ^ l> 1 ^ - x _a \

Cg - XcI ^esl 'UsG. - y r 's 1 cr - cn - ί ^y1 - y if! Y _c1 - usg \ <\ xsg - x _c1 \

U71P = 1,. . .

"Gg - smiling ^x ^e if \ y _c1 -> Ssg - 1 x _c1 dir =

32

tgphi = mn—

mx

If dir ^{= 0} , then totraPred.An.glel = a _if JntraPredAngle2 = _{i + l}

5 _G de ae ^{ -32, -26, -21, -17, -13, -9, -5, -2Л2Д9ДЗД7,21,26,32) _and ^ t ^ fci <cc _{i + 15 and the} numbers of the corresponding prediction modes lie in the range of 2-17.

If dir = 1, then IntraPredAnglel = -a _{t ^} IntraPredAngle2 - - _{i + 11} where а е {-32, -26, -21, -17, -13, -9, -5, -2Д2,5,9ДЗ, 17,21,26,32) _and

YU . <tgphi <a _{i + t 5 and the} numbers of the corresponding prediction modes lie in the range 18-34.

Thus, the proposed method allows you to create a list of candidate modes, the quality of the predictions of which are evaluated at the final stage according to the standard RDO procedure. List length

15 is limited from above to six positions. The formation of the list does not require the prediction procedure itself, which leads to a significant reduction in the amount of computation when performing spatial prediction of the pixel values of the encoded block.

twenty

25

Claims

Claim

A method for quickly selecting the spatial prediction mode in the HEVC coding system, which consists in constructing a shortened list of spatial prediction candidate modes with a length of not more than six positions, based on the analysis of the high-frequency horizontal and vertical components of the stationary wavelet transform of the Haar encoded image region, while the array of values of the high-frequency horizontal component Haar transform is obtained by subtracting from the value of each pixel encoded about the values of the pixel adjacent to the left, while the array of values of the high-frequency vertical component of the Haar transform is obtained by subtracting from the value of each pixel of the encoded region the values of the neighboring neighboring pixels, while the Planar mode is included in the list of candidate modes if all the absolute values of the horizontal and vertical high-frequency components are less quantization steps during encoding, while the DC mode is included in the list of candidate modes if the standard deviation of the horizontal values of vertical and high-frequency components is less than the quantization step, and two angular prediction modes corresponding to the mutual displacement of the points of intersection of the line of the minimum values of the gradient modulus passing through the maximum point of the array of vertical high-frequency components with vertical boundaries of the array of high-frequency components are included in the list of candidate modes, while two angular prediction modes corresponding to the mutual displacement of the points of intersection of the line of minimum values of the gradient modulus entent passing through the maximum point of the array of horizontal high-frequency components, with horizontal the boundaries of the array of high-frequency components are included in the list of candidate modes, and the final selection of the prediction mode for the encoded image area from the constructed list of candidate modes is based on standard RDO (rate-distortion optimization) estimates.