JP5423469B2 - Quantization apparatus, program and method, and moving picture encoding apparatus - Google Patents

Description

  The present invention relates to a quantization apparatus, a program and method, and a moving picture encoding apparatus. The present invention can be applied to processing of a moving image compression code by a moving image encoding technique such as H.264.

  H. H.264 (ITU-T Rec.H264 | ISO / IEC 14496-10) and other moving image information compression coding processing using moving image coding technology is performed for each processing unit obtained by dividing an input image. The prediction residual signal, which is the difference between the predicted image that has undergone compensated prediction and the input image, is quantized with a transform coefficient that has been subjected to orthogonal transform such as discrete cosine transform, and this is then entropy such as variable-length code. Encoding realizes highly efficient video compression.

  At this time, quantization noise is generated by the quantization process, and distortion occurs in the decoded image reproduced on the decoding side. On the other hand, by quantizing, the amount of information to be encoded is reduced, and highly efficient compression is realized.

  There is a rate distortion optimization technique as a technique for evaluating such a trade-off between distortion and rate (code amount) and using it for coding mode selection or the like.

  That is, for options such as a plurality of encoding modes, the distortion D of the decoded image obtained when the option is selected, the code amount R generated when the option is encoded, and the Lagrange multiplier λ In this method, encoding is performed with an optimum trade-off between rate and distortion by making a selection that minimizes the RD cost J (J = D + λR).

  Usually, the distortion D is evaluated by the sum of square errors between the decoded pixel value and the input pixel value.

  Non-Patent Document 1 discloses a method of using this rate distortion optimization technique for transform coefficient quantization.

  In the method of Non-Patent Document 1 (conventional RDOQ (Rate Distortion Optimized Quantization)) In the quantization process of the H.264 encoding process, when distortion D is evaluated as the sum of square errors of pixel values, the sum of square errors in the pixel value area is calculated as 2 in the transform coefficient area due to the nature of orthogonal transformation. Utilizing the fact that it can be approximated by the sum of multiplication errors, it is formulated as a method for minimizing the RD cost (which can be expressed as the following equation (1)) of a given block.

J (λ) = Σ _i Σ _j J (λ, c _ij , l _ij ) (1)
In the equation (1), J (λ, c _ij , l _ij ) is an RD cost when the transform coefficient c _ij is quantized to the coefficient level l _ij and encoded, as in the following equation (2) Can show.

J (λ, c _ij , l _ij ) = err (c _ij , l _ij )
+ Λbits (l _ij ) (2)
In the equation (2), the quantization error err (c _ij , l _ij ) can be expressed as the following equation (3).

err (c _ij , l _ij ) = N (QP% 6, i, j) (r _ij −c _ij ) ² (3)
In the equation (3), r _ij is a value obtained by dequantizing the coefficient level l _ij and can be expressed as the following equation (4).

r _ij = l _ij R (QP% 6, i, j) (4)
In the equations (3) and (4), N (QP% 6, i, j) and R (QP% 6, i, j) are respectively H. The normalization matrix and the inverse quantization scale matrix of the integer transform used in H.264, and QP is a quantization parameter. Note that QP / 6 and QP% 6 represent the quotient and remainder obtained by dividing QP by 6.

Also, in the expression (2), bits (l _ij ) is an encoding amount that is required when l _ij is encoded by CAVLC (Context-based Adaptive Length Coding) or CABAC (Context-based Adaptive Binary Coding). It is.

Further, in Non-Patent Document 1, an algorithm for selecting a candidate for minimizing the RD cost from a plurality of given c _ij to l _ij for the case where the entropy code used is CABAC and CAVLC. Is disclosed.

In the CABAC algorithm, candidates for | l _ij | are mainly selected from three types of 0, l _ij ^floor , and l _ij ^{ceil in} the following equations (5) to (8). In order from the high frequency side, in the CAVLC algorithm, the above-mentioned RD cost J (λ) is set to each coefficient component candidate for each coefficient component mainly in the order of the absolute value of the coefficient from the three ^types of 0, 1, and l _ij ^round. Each coefficient level is evaluated and the coefficient level that gives the minimum value is selected.

l _ij ^float = | c _ij | Q (QP% 6, i, j) / 2 ^{15 + QP / 6} (5)
^{_{l ij floor = floor (l ij}} float) ... (6)
l _ij ^ceil = l _ij ^floor +1 (7)
l _ij ^round =
floor {(| c _ij | Q (QP% 6, i, j) +0.5) / 2 ^{15 + QP / 6} } (8)

By Marta Karczewicz, Yan Ye, Peisong Chen, “Rate Distribution Optimized Quantization”, [Online], INTERNET, [Search February 10, 2010], <URL: http://ftp3.itu.int/av-arch/ jvt-site / 2008_04_Geneva / JVT-AA026.zip>

  However, Non-Patent Document 1 (conventional RDOQ) has a problem that the amount of calculation for performing the optimization process of quantization is very large, and an optimal quantization result cannot be obtained with sufficient processing performance. .

That is, for example, in order to obtain the quantization error err (c _ij , l _ij ), two multiplications, one multiplication with λ, and at least three multiplications evaluate the RD cost of a certain coefficient level candidate. If three candidates are evaluated, about nine multiplications are required for each coefficient.

In addition, for example, in the case of CAVLC, l _ij ^round is evaluated, but _ij ^floor or l _ij ^ceil, which is different from l _ij ^round , is not evaluated, so the RD cost with a relatively high possibility is high. There is a problem that the coefficient level to be decreased may not be selected. For example, even if the number of selection candidates is simply increased, the processing amount may be increased.

  Therefore, a quantization device, a program and a method, and a moving image encoding device that can reduce the amount of processing in the quantization processing related to moving image encoding are desired.

  The first aspect of the present invention is a quantization apparatus that quantizes a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image in the moving image encoding process. (1) scaling means for performing scaling for quantization on the transform coefficient; and (2) a first coefficient level compensation obtained by rounding the absolute value of the scaled transform coefficient into an integer by rounding off the decimal point. , A coefficient level candidate determining means for determining a second coefficient level candidate having an absolute value one larger than that of the first coefficient level candidate, (3) when the first coefficient level compensation is selected, and the second When a coefficient level candidate is selected, a scaled distortion change amount corresponding to the difference in quantization error and a code amount change amount necessary for entropy coding are obtained, and the obtained distortion change is obtained. The amount of change in the rate distortion cost between the case where the first coefficient level compensation is applied and the case where the second coefficient level compensation is applied is calculated using the amount and the change amount of the code amount. Rate distortion cost change amount calculation means; and (4) depending on the calculation result of the rate distortion cost change amount calculation means, either the first coefficient level candidate or the second coefficient level candidate is converted to the conversion coefficient. And coefficient level selection means for selecting a coefficient level to be applied to the quantization.

  The second aspect of the present invention is a quantization program for quantizing a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image in the moving image encoding process. (1) scaling means for performing scaling for quantization on the transform coefficient; and (2) a first coefficient level obtained by converting the absolute value of the scaled transform coefficient into an integer by rounding off a decimal point. Compensation, coefficient level candidate determination means for determining a second coefficient level candidate whose absolute value is one greater than that of the first coefficient level candidate, and (3) when the first coefficient level compensation is selected. When the second coefficient level candidate is selected, the scaled distortion change amount corresponding to the difference in quantization error and the change in the code amount required for entropy coding are used. A rate between the case where the first coefficient level compensation is applied and the case where the second coefficient level compensation is applied using the obtained distortion change amount and code amount change amount Rate distortion cost change amount calculating means for calculating the distortion cost change amount; and (4) the first coefficient level candidate or the second coefficient level candidate according to the calculation result of the rate distortion cost change amount calculating means. Any one of the above is functioned as coefficient level selection means for selecting a coefficient level to be applied to quantization of the transform coefficient.

A third aspect of the present invention is a quantization method for quantizing a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image in the moving image encoding process. (1) a scaling step for performing scaling for quantization on the transform coefficient; and (2) a first coefficient level compensation obtained by converting the absolute value of the scaled transform coefficient into an integer by rounding down the decimal point. The coefficient level candidate determination step for determining a second coefficient level candidate whose absolute value is one greater than that of the first coefficient level candidate; (3) when the first coefficient level candidate is selected; When a coefficient level candidate of 2 is selected, the scaled distortion change amount corresponding to the difference in quantization error and the change amount of the code amount necessary for entropy coding are obtained and obtained. Using the change amount and the change amount of the code amount, the change amount of the rate distortion cost between the case where the first coefficient level compensation is applied and the case where the second coefficient level compensation is applied is calculated. A rate distortion cost change amount calculating step, and (4) converting the first coefficient level candidate or the second coefficient level candidate into the conversion according to the calculation result of the rate distortion cost change amount calculating step. And a coefficient level selection step of selecting as a coefficient level to be applied to coefficient quantization.

  According to a fourth aspect of the present invention, there is provided a quantization device that quantizes a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image in the moving image encoding device. The quantization apparatus according to the first aspect of the present invention is applied as the quantization apparatus.

  According to the present invention, it is possible to reduce the amount of processing in the quantization processing related to moving image coding.

It is the block diagram shown about the functional structure of the quantization part (quantization apparatus) which concerns on 1st Embodiment. It is the block diagram shown about the functional structure of the moving image encoder which concerns on 1st Embodiment. It is the flowchart (1) shown about operation | movement of the quantization part (quantization apparatus) which concerns on 1st Embodiment. It is the flowchart (2) shown about operation | movement of the quantization part (quantization apparatus) which concerns on 1st Embodiment. It is the flowchart shown about operation | movement of the quantization part (quantization apparatus) which concerns on 2nd Embodiment. It is the flowchart (1) shown about operation | movement of the quantization part (quantization apparatus) which concerns on 3rd Embodiment. It is the flowchart (2) shown about operation | movement of the quantization part (quantization apparatus) which concerns on 3rd Embodiment. It is the flowchart shown about operation | movement of the quantization part (quantization apparatus) which concerns on 4th Embodiment. It is the block diagram shown about the functional structure of the quantization part (quantization apparatus) which concerns on 5th Embodiment.

(A) First Embodiment Hereinafter, a first embodiment of a quantization apparatus, a program and a method, and a moving picture coding apparatus according to the present invention will be described in detail with reference to the drawings. Note that the quantization apparatus according to the first embodiment will be described as a quantization unit.

(A-1) Configuration of First Embodiment (A-1-1) Overall Configuration of Video Encoding Device FIG. 2 is a block diagram showing the overall configuration of the video encoding device 10 of this embodiment. is there.

  The moving image encoding apparatus 10 includes a prediction difference signal generator 11, a conversion unit 12, a quantization unit 13, an entropy encoding unit 14, an inverse quantization / inverse conversion unit 15, a decoded image generation unit 16, a frame buffer 17, and a prediction. A portion 18 is provided. Note that components other than the quantization unit 13 are, for example, the existing H.264. An apparatus similar to the moving picture encoding apparatus compliant with H.264 can be applied.

  The moving image encoding apparatus 10 is an apparatus having a program execution configuration such as a CPU, ROM, RAM, EEPROM, and hard disk (not limited to one, but may be configured to be capable of distributed processing of a plurality of units. )) May be constructed by installing a moving picture encoding program (including the quantization program of the embodiment) or the like, and even in that case, it can be functionally shown as in FIG. .

  The prediction difference signal generator 11 obtains a prediction error signal from the supplied image data of the original image and the prediction image supplied from the prediction unit 18 and supplies the prediction error signal to the conversion unit 12.

  The transform unit 12 performs orthogonal transform processing on the prediction error signal, obtains a prediction error signal converted into a transform coefficient such as a frequency component, and supplies the prediction error signal to the quantization unit 13.

  The quantization unit 13 quantizes the prediction error signal converted into a conversion coefficient such as a frequency component, obtains a quantized prediction error signal, and supplies the quantized prediction error signal to the entropy encoding unit 14 and the inverse quantization / inverse conversion unit 15. To do.

  The entropy encoding unit 14 performs entropy encoding (for example, variable length encoding) on the quantized prediction error signal, and generates stream data together with other encoding information. The data of this stream becomes the output of the moving image encoding device 10.

  The inverse quantization / inverse transform unit 15 performs inverse quantization and inverse orthogonal transform on the quantized prediction error signal, obtains a prediction error signal with a quantization error, and supplies it to the decoded image generation unit 16.

  The decoded image generation unit 16 obtains a decoded image from the prediction error signal accompanied with the quantization error and the prediction image.

  The frame buffer 17 stores a decoded image that has already been encoded and locally decoded as a reference image. H. In the case of the H.264 / AVC standard, an image to be referred to in motion compensation is not limited to an image immediately before an image to be currently encoded, and images at a plurality of times can be reference images.

  The prediction unit 18 obtains prediction information such as a motion vector from image data at a target time to be encoded and a reference image at a time different from the target time, generates predicted image data, and generates a prediction difference signal. To the device 11 and the decoded image generator 16. H. In the case of H.264, a prediction image based on intra prediction using a decoded image that has already been encoded at the target time and is locally decoded may be supplied.

(A-1-2) Processing Outline of Quantization Unit Next, an outline of processing by the quantization unit 13 will be described.

The quantization unit 13 performs scaling for quantization on a transform coefficient (for example, a transform coefficient that has been subjected to orthogonal transform such as discrete cosine transform), and converts the absolute value of the scaled value c _i to a _i , Let the sign of c _{i be} s _i . Then, a t _i that integer by truncating the decimal point of a _i, at least two t i _+1, a candidate of the absolute value u _i of the coefficient levels l _i quantized. Furthermore, a scaled distortion change amount ΔE corresponding to a difference in quantization error between when t _i is selected as u _i and when t _i +1 is selected (hereinafter simply referred to as a distortion change amount). To describe). Furthermore, as u _i, needed to entropy encode If you choose t _i or t _i +1, obtaining the code amount of variation △ R.

Also, the amount of change in the scaled RD cost from the selection of t _i to the selection of t _i +1 using the Lagrange multiplier μ scaled for comparison in the coefficient level region of the Lagrangian multiplier λ (hereinafter referred to as “r _i”). Simply described as the amount of change in RD cost) by the following equation (9). Then, it is configured to select t _i or t _i +1 as the absolute value u _i of the quantized coefficient level l _i with respect to c _i scaled in the coefficient level region based on the positive / negative determination in the obtained result. . For simplification of explanation, the subscript i represents each coefficient component. For example, in the case of 4 × 4 conversion, 16 coefficient components are crossed. In the following description, it is assumed that numbers are assigned in the order of scanning such as Zig-Zag scanning unless otherwise specified.

ΔJ = ΔE + μΔR (9)
Here, the meanings of ΔJ, ΔE, μ, and ΔR will be described.

H. When the inverse quantization and inverse transform used in the H.264 decoding process are regarded as real number operations ignoring truncation of integer operations in the calculation process, the value c _i of the input coefficient level region and the pixel value to be output If the quantization parameter is QP between the region values d _i and q = (2 ^1/6 ) ^(QP−4) (where 2 ^1/6 is the sixth root of 2 as a real number), The linear transformation is defined so that the relationship of the following expression (10) is substantially satisfied.

Σ _i d _i ² = Σ _i q ² c _i ² (10)
That is, it can be regarded as an approximate transformation of orthonormal transformation multiplied by the quantization step size q.

Therefore, when the value c _i that scale factor level region transform coefficient is quantized in the coefficient level l _i with integer values, the quantization error e occurring in the region of the pixel values of the decoded image (c _i, l _i) Can be approximated by the following equation (11).

e (c _i , l _i ) = q ² (c _i −l _i ) ² (11)
Since c _i and l _i have the same sign and are squared, the sign may be ignored and the evaluation may be performed using only the absolute value. When each absolute value is represented by a _i and u _i , the following (12 )

^{_{e (a i, u i)}} = q 2 (a i -u i) 2 ... (12)
The amount of change in quantization error when t _i and t _i +1 are selected as u _i is expressed by the following equation (13).

Δe (a _i , t _i ) = e (a _i , t _i +1) −e (a _i , t _i )
= 2q ² (1 / 2− (a _i −t _i )) (13)
If the whole is scaled to 1 / (2q ² ) and evaluated, it can be evaluated without multiplication as in the following equation (14).

ΔE (a _i , t _i ) = Δe (a _i , t _i ) / (2q ² )
_{= 1 / 2- (a i -t} i) ... (14)
At this time, if μ is set as in the following equation (15), the value ΔJ obtained by scaling the change amount of the RD cost J by 1 / (2q ² ) can be evaluated by the above equation (9).

μ = λ / (2q ² ) (15)
At this time, since ΔR is the amount of change in the code amount, when t _i or t _i +1 is selected as u _i , the amount of change in the code amount of only the syntax element whose code amount changes may be obtained.

△ If J is negative (△ J <0), it represents the person to select a _t i +1 as _{u i} is RD cost is reduced, △ by determining the sign of J, _{t i} and _t i +1 Can be selected.

Here, H. When an H.264 encoder is taken as an example, an H.264 code shown in the following equation (16) is used for a residual signal that is a difference between an input original image and a predicted image obtained by motion compensation or the like. A coefficient level that is optimal in terms of rate distortion as a coefficient level obtained by applying a transform coefficient obtained by applying H.264 integer transform to a two-dimensional input and quantizing using a predetermined quantization parameter QP Select a value and output it. The coefficient level value obtained by the quantization process is entropy-encoded by an entropy encoder such as CABAC or CAVLC after selecting an encoding mode or the like, and is output as a moving image stream. Further, the obtained coefficient level is inversely quantized / inverted and added to the predicted image to reproduce the decoded image, and is used to generate the next predicted image.

  Hereinafter, the quantization unit 13 will be described as a process in which an input is a transform coefficient and an output is a coefficient level.

  H. In H.264, in addition to the above 4 × 4 transform, there are 8 × 8 transform and Hadamard transform for a block in which DC coefficients are collected. However, since it can be similarly configured, 4 × 4 transform is taken as an example below. explain.

(A-1-3) Internal Configuration of Quantization Unit FIG. 1 is a block diagram showing a functional configuration of the quantization unit 13.

  The quantization unit 13 includes a scale calculation unit 131, a distortion change amount calculation unit 132, a rate change amount calculation unit 133, and a coefficient level selection unit 134.

  The quantization unit 13 (quantization device) is a device having a program execution configuration such as a CPU, ROM, RAM, EEPROM, hard disk, etc. (not limited to one, but capable of distributed processing of a plurality of devices). May be constructed by installing the quantization program or the like of the embodiment, and even in that case, it can be functionally shown as in FIG. In this embodiment, the quantization unit 13 is illustrated as being mounted on the moving image encoding device 10, but may be configured as a single device (quantization device). When the moving image encoding apparatus 10 is configured by installing a moving image encoding program in the information processing apparatus described above, the quantization unit 13 is a quantization program that is a part of the moving image encoding program. You may make it build as.

In the quantization unit 13, the scale calculation unit 131 performs scaling according to the quantization parameter QP on the transform coefficient obtained by integer transform. That is, each element (hereinafter also referred to as “coefficient component”) of the scaled coefficient c _jk is obtained by the following expression (17) with respect to the conversion coefficient g _jk .

c _jk = g _jk Q (QP% 6, j, k) / 2 ^{15 + QP / 6} (17)
Here, Q (QP% 6, j, k) is a scale matrix at the time of quantization, and QP / 6 and QP% 6 are the quotient obtained by dividing QP by 6, respectively, and the division by 215 ^{+ QP / 6} is Division as a real number.

Note that this division can be performed by a right shift operation using a number obtained by subtracting the number of bits after the decimal point from 15 + QP / 6 if c _jk is expressed by a fixed decimal number with appropriate accuracy. For example, when c _jk is expressed by a fixed decimal having 11 bits after the decimal point, it is a 4 + QP / 6 bit right shift.

  In conversions other than 4 × 4, the scale matrix, the shift amount, and the like are different, but the same processing is performed to scale the values in the conversion coefficient region to values that are not integerized in the coefficient level region.

Here, for simplification of the following description, the scaled coefficient c _ij will be described again as c _i in the Zig-Zag scan order, and its absolute value as shown in the following equation (18). A _i , and the code as s _i as shown in equation (19). However, i = 0,..., M, where M is the number of coefficient components of the target coefficient block−1.

a _i = | c _i | (18)
s _i = sign (c _i ) (19)
In the following description, operations for these values are described as real number operations, but can be processed by an integer arithmetic unit using a fixed decimal representation with appropriate precision. The right shift at the time of scaling may be performed on the absolute value of g _jk .

If the integer part of the absolute value a _i (the maximum integer not exceeding a _i ) is set to t _i , t _i is expressed as in the following equation (20).

t _i = floor (a _i ) (20)
Then, as a candidate of the absolute value _{u i} of the coefficient levels _{l i} relative to the scale coefficients _{c i,} and _{t i,} and at least two candidates for _t i +1 (△ skipping evaluation of R, a _{t i} Including the case of selection).

  The distortion change amount calculation unit 132 calculates the following equation (21).

ΔE (a _i , t _i ) = 1 / 2− (a _i −t _i ) (21)
The rate change amount calculation unit 133 obtains a change amount ΔR of the code amount that will be generated for both candidates by respective entropy encoders such as CABAC and CAVLC. The processing of the rate change amount calculation unit 133 will be described in detail in the operation description to be described later.

The coefficient level selection unit 134 uses the scaled Lagrangian multiplier μ that has been calculated in advance to determine the optimum value based on the determination corresponding to the determination of the sign of the RD cost change evaluation formula (the above-described formula (9)). The coefficient level value l _i is output by selecting a suitable coefficient level.

(A-2) Operation of the First Embodiment Next, the video encoding device 10 of the first embodiment having the above configuration will be described focusing on the operation of the quantization unit 13.

  In the first embodiment, when the entropy encoding performed by the entropy encoding unit 14 is CABAC, the processing in the quantization unit 13 is performed in the same processing procedure (decision order of coefficient components) as in Non-Patent Document 1. An example of selecting coefficient levels of individual components of the obtained coefficient block will be described. Since the entropy coding by CABAC has a syntax that represents the final non-zero coefficient in the scan order of the coefficient component in the transform coefficient, in the first embodiment, a quantization process is performed in consideration of the characteristics of the CABAC. Is called.

  FIG. 3 is a diagram illustrating a processing procedure of a quantization process in the quantization unit 13 of the first embodiment.

First, the quantization unit 13 obtains a scaled coefficient c _i obtained by scaling the input transform coefficient in accordance with the quantization parameter QP (S101). Here, i = 0,..., M is assumed to be a scan order such as a Zig-Zag scan.

Then, the quantization unit 13, the absolute value _{a i} of _{c i,} as Hoho absolute value _{u i} of the coefficient levels _{l i, t i} is calculated by the above expression (20) (S102).

Then, the quantization unit 13 obtains the distortion change amount ΔE _i by the above equation (21). Here, for _i where ΔE _i ≧ 0, u _i = t _i, and in the subsequent processing, t _i +1 is not a candidate.

  Next, the quantization unit 13 determines the position k of the last coefficient (the largest number) in the scan order among the transform coefficients for selecting the non-zero coefficient level (S103).

  FIG. 4 is a flowchart showing details of the final non-zero coefficient position determination process in step S103.

In FIG. 4, i ₀ ,... I ₁ are considered as k candidates. Here, i ₁ is the maximum value of i for which a _i > 0.5, and i ₀ is the maximum value of i for which a _i > 1 (-1 if not).

In the following description, bits () is an approximate value of the number of bits necessary to encode information in parentheses with CABAC (calculated using a probability state variable in the context of the syntax to be encoded, an arithmetic code. Therefore, the number of bits is not necessarily an integer). Also, coded is coded_block_flag, last _i is last_significant_coeff_flag, and u _i is a syntax related to coefficient level l _i combined with associated code s _i , significant_coeff_flag_

First, it is assumed that i ₀ ,..., I ₁ are set as candidate ranges (S201).

  Next, each variable is initialized by the following formulas (22) to (25).

k = i ₀ (22)
e = 0 (23)
r = bits (coded = 1) −bits (coded = 0)
(When i ₀ == − 1) (24)
r = bits (last _k = 0) −bits (last _k = 1)
(When i ₀ ≧ 0) (25)
Next, updating of k, e, r is repeated within the candidate range (i = i ₀ +1,..., I ₁ ) (S203, S204).

  In steps S203 and S204, first, p and b are obtained by the following equations (26) and (27), respectively. Then, k, e, and r are updated according to the results of p and b.

p = ΔE _i + μ (bits (u _i = 1, last _i = 0)
-Bits (u _i = 0)) (26)
b = e + ΔE _i + μ (r + bits (u _i = 1, last _i = 1) (27)
[For b <0]
Update to k = i.

  In the case of p <0, e and r are obtained as in the following equations (28) and (29).

e = 0 (28)
r = bits (last _k = 0) −bits (last _k = 1) (29)
In the case of p ≧ 0, e and r are obtained as in the following equations (30) and (31).

e = −ΔE _i (30)
r = bits (u _i = 0) −bits (u _i = 1, last _i = 0) (31)
[For b ≧ 0]
In the case of p <0, e and r are obtained as in the following equations (32) and (33).

e = e + ΔE _i (32)
r = r + bits (u _i = 1, last _i = 0) (33)
In the case of p ≧ 0, r is obtained as in the following equation (34).

r = r + bits (u _i = 0) (34)
As described above, after k, e, and r are updated in steps S203 and S204, the obtained k is set as the final non-zero coefficient position (S205). However, when k = −1, the target block has no coefficient.

  Here, the variables used in the flowchart shown in FIG. 4 will be described.

  “K” is a provisional final non-zero coefficient position in the above process (k = −1 indicates no non-zero coefficient).

  “E” is a value obtained by accumulating the amount of change in distortion from the case where k is the final non-zero coefficient position to the case where k is not the final position up to i.

  “R” is a value obtained by accumulating the amount of change in the code amount from when k is the final non-zero coefficient position to when i is not the final position up to i.

  The meaning of b <0 indicates that the RD cost is reduced when the position of i is the final non-zero coefficient position than when k is set.

  The meaning of p <0 means that if it is assumed that the position in the scanning order larger than i is the final non-zero coefficient position, selecting 1 as the absolute value of the coefficient level at the position of i reduces RD cost. Represents what to do.

Next, the coefficient level selection unit 134 uses the k obtained by the processing of FIG. .., 0, coefficient levels are selected (S104). That is, the k in the reverse order of scan, for each i, the code amount of the change amount △ R _i, the following (35) determined as equation below (36) the variation of the RD cost in the expression of the code Depending on the judgment, the coefficient level is selected from t _i or t _i +1. That is, t _i +1 is selected when ΔJ _i <0. As described above, for i for which ΔE _i ≧ 0, u _i = t _i and this determination is skipped.

ΔR _i = bits (u _i = t _i +1) −bits (u _i = t _i ) (35)
ΔJ _i = ΔE _i + μΔR _i (36)
Here, the change amount ΔR _i of the code amount is further efficiently obtained for each of the following four cases depending on the magnitude of t _i depending on the coefficient level expression method in CABAC. Can do. The code amount by the CABAC arithmetic code depends on the probability state variable, and it is better to evaluate the state variable repeatedly used while performing state transition, but it may be approximated without state transition.

[If t _i > 14]
In this case, ΔR _i does not depend on the probability state variable, only the difference of the suffix part length of coeff_abs_level_minus1 is obtained, and ΔR _i is 0 or 2.

[When 2 ≦ t _i ≦ 14]
In this case, ΔR _i depends on only one probability state variable related to coeff_abs_level_minus1, and a table of ΔR _i can be created in advance with t _i having 13 patterns and 128 probability state variables.

  Also, in CABAC, this one probability state variable is repeatedly used while performing state transition during the encoding of one coeff_abs_level_minus1, so that the code amount difference after state transition can also be tabulated.

[When t _i = 1]
In this case, ΔR _i depends on two probability state variables related to coeff_abs_level_minus1, but it is only necessary to obtain the code amount difference between the two bins.

[t _i = 0]
In this case, ΔR _i is related only to the first bin of the significant_coeff_flag and coeff_abs_level_minus1, and can be obtained from two probability state variables related thereto.

  With the above processing, it is possible to select a coefficient level that is optimal in terms of rate distortion equivalent to the conventional RDOQ CABAC algorithm with a smaller amount of computation and processing.

(A-3) Effect of First Embodiment According to the first embodiment, the coefficient level that is optimal in terms of rate distortion can be selected with less processing as in the RDOQ CABAC algorithm described in Non-Patent Document 1. It can be performed in quantity (calculation amount).

(B) Second Embodiment Hereinafter, a second embodiment of a quantization apparatus, a program and a method, and a moving picture coding apparatus according to the present invention will be described in detail with reference to the drawings. Note that the quantization apparatus according to the second embodiment will be described as a quantization unit.

(B-1) Configuration and Operation of Second Embodiment The quantization unit and the moving image encoding device of the second embodiment can also be shown in FIG.

  In the second embodiment, the entropy encoding unit 14 supports CAVLC, and the quantization unit 13 performs processing procedures similar to those of Non-Patent Document 1 (decision order of coefficient components) for given coefficient blocks. This is different from the first embodiment in that the coefficient level of each component is selected. Since other configurations are the same as those in the first embodiment, detailed description thereof will be omitted, and in the following, the second embodiment will be described focusing on differences from the first embodiment.

  Since entropy encoding processing by CAVLC includes run-length encoding of zero coefficients, in the second embodiment, quantization is performed in consideration of this characteristic of CAVLC.

  FIG. 5 is a flowchart illustrating the processing of the quantization unit 13 according to the second embodiment.

First, the quantization unit 13 obtains a scaled coefficient c _i obtained by scaling the input transform coefficient according to the quantization parameter QP (S301). However, i = 0,..., M are numbered in the order in which the scaled coefficients c _jk are sorted in ascending order of their absolute values, unlike the first embodiment (S302). That is, c ₀ is a coefficient having the minimum absolute value.

Next, the quantization unit 13 initializes the absolute value a _i of c _i as an initial value of the absolute value u _i of the coefficient level l _{i by} the following equation (37) (S303). Here, the selection of the initial value of the coefficient level is, for example, in the case of f = 1/3 for an intra picture and f = 1/6 for an inter picture, as in the conventional RDOQ, as f in Expression (37). Other various simple quantization methods may be used.

u _i = floor (a _i + f) (37)
Also, the quantization unit 13, as an absolute value _{u i} candidates coefficient levels _{l i,} obtains a _{t i} by the following (38) to obtain the strain variation △ _{E i} according to the following (39).

t _i = floor (a _i ) (38)
ΔE _i = ΔE (a _i , t _i ) = 1 / 2− (a _i −t _i ) (39)
Here, for _i where ΔE _i ≧ 0, u _i = t _i, and in the subsequent processing, t _i +1 is not a candidate.

With the above preparation, the quantizing unit 13 selects coefficient levels l _i in the order i = 0,..., M in which the absolute values of the scaled coefficients c _i are small (S304). That is, in order from c _{0 having the} smallest absolute value, for each i, the change amount ΔR _i of the code amount when t _i is selected as u _i and when t _i +1 is selected, and the following equation (40) is obtained: The coefficient level is selected from t _i or t _i +1 by determining the sign of the amount of change in the RD cost shown in FIG. That is, t _i +1 is selected when ΔJ _i <0. The coefficient level l _i is sequentially updated using the selected u _i .

ΔJ _i = ΔE _i + μΔR _i (40)
As described above, for i for which ΔE _i ≧ 0, u _i = t _i and this determination is skipped.

When calculating the amount of change ΔR _i of the code amount, if necessary, the initial value is maintained, or if it is updated by the above processing, it is calculated with reference to the updated coefficient level l _i .

Here, regarding the amount of change ΔR _i of the code amount, only the syntax element having a change in the code amount between the case where t _i is selected and the case where t _i +1 is selected may be considered. That is, the syntax elements related to the calculation process of ΔR _i can be efficiently narrowed for each of the following cases.

[First requirement]
When t _i > 1, it does not affect the coeff-token and can be ignored.

[Second requirement]
When t _i > 0, total_zeros and run_before are not affected and can be ignored.

[Third requirement]
If there is no effect on the update of suffixLength, trailing_ones_sign_flag, level_prefix, level_suffix, etc. relating to coefficient levels other than the i-th coefficient can be ignored. In addition, it is good also as a process which ignores the influence on suffixLength and always ignores the syntax regarding coefficient levels other than the i-th.

[Fourth requirement]
If there is no change in level_prefix, ΔR _i = 0.

(B-2) Effects of Second Embodiment According to the second embodiment, the following effects can be achieved.

  Selection of the optimum coefficient level in terms of rate distortion can be performed with a smaller amount of calculation and processing amount as in the conventional RDOQ CAVLC algorithm described in Non-Patent Document 1.

  In addition, since the conventional RDOQ CAVLC algorithm described in Non-Patent Document 1 is not evaluated, it is likely to be optimal in terms of rate distortion. Efficiency can be improved.

(C) Third Embodiment Hereinafter, a third embodiment of a quantization device, a program and a method, and a moving image coding device according to the present invention will be described in detail with reference to the drawings. Note that the quantization apparatus according to the third embodiment will be described as a quantization unit.

(C-1) Configuration and Operation of Third Embodiment A quantization unit and a moving image encoding apparatus of the third embodiment can also be shown in FIG.

  In the third embodiment, the entropy encoding unit 14 supports CABAC, and the quantization unit 13 is provided with a coefficient block in a processing procedure (decision order of coefficient components) different from that of the first embodiment. This is different from the first embodiment in that the coefficient level of each component is selected. Since the other configuration is the same as that of the first embodiment, a detailed description thereof will be omitted. In the following, the third embodiment will be described focusing on differences from the first embodiment.

  FIG. 6 is a flowchart illustrating the processing of the quantization unit 13 according to the third embodiment.

First, the quantization unit 13 obtains a scaled coefficient c _i obtained by scaling the input transform coefficient according to the quantization parameter QP (S401). (Where i = 0,..., M is the scan order such as Zig-Zag scan)
Then, the quantization unit 13, the absolute value _{a i} of _{c i,} as a candidate of the absolute value _{u i} of the coefficient levels _{l i,} obtains a _{t i} by the following equation (41), further, the following (41) The strain change amount ΔE _i is obtained from the equation.

t _i = floor (a _i ) (41)
ΔE _i = ΔE (a _i , t _i ) = 1 / 2− (a _i −t _i ) (42)
Here, for _i where ΔE _i ≧ 0, u _i = t _i, and in the subsequent processing, t _i +1 is not a candidate.

  Next, the position k of the last coefficient (the largest number) in the scan order among the transform coefficients for selecting the non-zero coefficient level is determined (S402, S403).

The difference from the first embodiment (FIG. 4 described above) is that processing is performed in the reverse order of scanning. In other words, in the first embodiment, i ₀ ,..., I ₁ that are candidate ranges of k are processed in the forward direction to determine k, but in the third embodiment, i = M In the reverse order of scanning, the position of i ₁ is obtained (S402), and the position of k is searched for up to the position corresponding to i ₀ (S403).

The quantization unit 13 considers i ₁ ,..., I ₀ as candidates for k. However, i ₁ is the maximum value of i with a _i > 0.5, and i ₀ is the maximum value of i with a _i > 1 (-1 if not). If there is no such i ₁ , the target block has no coefficient, and the following processing is not performed.

That is, the quantization unit 13 sets the position of i ₁ as the process of step S402 described above, but if there is no i where a _i > 0.5, the target block is assumed to have no coefficient, and the following step S403 is performed. No processing is performed.

  FIG. 7 is a flowchart showing details of the processing in step S403 described above in the quantization unit 13 of the third embodiment.

The quantization unit 13 determines k while searching for the position of i ₀ by the process shown in FIG. However, in the following, the meaning of bits () is the same as in the first embodiment.

First, assuming that the initial value of k is k = i ₁ and a _k > 1 at this time, i ₀ = i ₁ , so the following search processing of k and i ₀ is not performed, and i ₀ or less Only the process (S404) is performed.

  Then, the quantization unit 13 initializes each variable (k, e, r) using the following equations (43), (44), and (45) (S501).

k = i ₁ (43)
e = ΔE _k (44)
r = bits (u _k = 1, last _k = 1) (45)
Next, the quantization unit 13 obtains ΔE _i , p, b for i = i ₁ −1,..., 0, respectively, and updates u _i , k, e, r according to the result, i Repeat until ₀ is found (S502, S503). However, if t _i > 1, i is assumed to be i ₀ = i and i ₀ is found (S504), and the process proceeds to step S507 described later.

Next, the case classification when updating u _i , k, e, and r according to the results of ΔE _i , p, and b in steps S502 and S503 described above will be described.

[If △ E _i > 0]
u _i is updated by the following equation (46), r is updated by the following equation (47), and the process proceeds to the next i.

u _i = 0 (46)
r = r + bits (u _i = 0) (47)
[If △ E _i ≦ 0]
In this case, first, p is set as (48) below, and processing corresponding to the result of p and b described later is performed.

p = ΔE _i + μ (bits (u _i = 1, last _i = 0)
−bits (u _i = 0)) (48)
[If ΔE _i ≦ 0 and p <0]
First, u _i is updated by the following equation (49), and b is assumed to be the following equation (50).

u _i = 1 (49)
b = e + μ (r + bits (last _i = 0)
-Bits (last _i = 1)) (50)
When b> 0, k, e, and r are updated by the following expressions (51), (52), and (53), respectively.

k = i (51)
e = ΔE _k (52)
r = bits (u _k = 1, last _k = 1) (53)
On the other hand, when b <= 0, e and r are updated by the following equations (54) and (55), respectively.

e = e + ΔE _i (54)
r = r + bits (u _i = 1, last _i = 0) (55)
[If ΔE _i ≦ 0 and p ≧ 0]
First, u _i is updated by the following equation (56), and b is set by the following equation (57).

u _i = 0 (56)
b = e + ΔE _i + μ (r + bits (u _i = 0)
_{-Bits (u k = 1, last} i = 1)) ... (57)
When b> 0, k, e, and r are updated by the following equations (58), (59), and (60), respectively.

k = i (58)
e = ΔE _k (59)
r = bits (u _k = 1, last _k = 1) (60)
On the other hand, when b <= 0, r is updated by the following (61).

r = r + bits (u _i = 0) (61)
As described above, when the search for i ₀ is performed by repeating steps S502 and S503, the quantization unit 13 determines the processing result (S504), and when i ₀ is found (t _i > in the case where 1 is found), it carried out the processing of step S507 to be described later, in a case where i ₀ is not found (t _i> If one is not found), the process of step S505 described below is performed.

In step S502, S503 described above, _{if i 0} is determined to not be found is the quantization unit _13, for _i 0, b, respectively, the following equation (62), placed as (63) below , B is determined (S505).

i ₀ = -1 (62)
b = e + μ (r + bits (coded = 1)-bits (coded = 1)) (63)
If b> 0, the target block has no coefficient (S506), and the process of FIG. 7 ends. Otherwise, the process proceeds to step S508, which will be described later.

In the above step S505, if b> 0 is not satisfied, k and i ₁ are further compared, and if k <i ₁ , the range of i = k + 1,..., I ₁ is cleared to u _i = 0. (S508), the process of FIG.

On the other hand, in step S502, S503 described above, when it is determined that _{i 0} is found, b at as follows (64) formula, when a b> 0 is updated with k = i (S507 On the other hand, if b ≦ 0, k is not updated (S507). That is, when b ≦ 0, the position larger than i ₀ is still indicated. Thereafter, regardless of the determination of b, if k <i ₁ , the range of i = k + 1,..., I ₁ is cleared to u _i = 0 (S508), and the process of FIG.

b = e + μ (r + bits (last _i = 0)
-Bits (last _i = 1) (64)
Next, variables used in the process of FIG. 7 will be described.

  k is a provisional final non-zero coefficient position in the above process (k = −1 indicates no non-zero coefficient).

  e is a value obtained by accumulating the amount of change in distortion from the case where k is not the final non-zero coefficient position to the case where k is the final position up to i.

  r is a value obtained by accumulating the amount of change in the code amount from the case where k is not the final non-zero coefficient position to the case where k is the final position up to i.

  The meaning of b> 0 indicates that the RD cost is reduced when the position of i is the final non-zero coefficient position than when k is set.

  The meaning of p <0 is that if it is assumed that the position k in the scanning order larger than i is the final non-zero coefficient position, it is more RD cost to select 1 than 0 as the absolute value of the coefficient level at the position of i. Represents a decrease.

Next, the process of step S403 described above (the processing in FIG. 7) is completed, the quantization unit 13, _{if i 0} is found, i _{= i} 0. ..., 0, coefficient levels are selected. That, _{i 0} (e.g., the search stop position) in order from, for each i, the code amount of variation △ the _{R i,} the following (65) as a formula, the following (66) of the RD costs of the variation in the expression The coefficient level is selected from t _i or t _i +1 depending on the sign. That is, t _i +1 is selected when ΔJ _i <0.

ΔR _i = bits (u _i = t _i +1) −bits (u _i = t _i ) (65)
ΔJ _i = ΔE _i + μΔR _i (66)
As described above, for i is △ E _i ≧ 0, and u i ₌ t _i, may be skip this determination.

Also, the code amount change ΔR _i can be efficiently obtained by performing calculation processing only on bins having a change in the code amount, as in the first embodiment.

(C-2) Effects of Third Embodiment According to the third embodiment, the following effects can be achieved in addition to the effects of the first embodiment.

In the first embodiment, the coefficients need to be scanned a plurality of times in the forward direction and in the reverse direction. In the third embodiment, until i ₁ is found in i = M only in the reverse direction, until i _{0 is} found. , It is only necessary to scan up to i = 0, so that the processing is simplified, the processing amount is reduced, and the memory access can be reduced.

(D) Fourth Embodiment Hereinafter, a fourth embodiment of a quantization device, a program and a method, and a moving image coding device according to the present invention will be described in detail with reference to the drawings. Note that the quantization apparatus according to the fourth embodiment will be described as a quantization unit.

(D-1) Configuration and Operation of Fourth Embodiment A quantization unit and a moving image encoding device of the fourth embodiment can also be shown in FIG.

  In the fourth embodiment, the entropy encoding unit 14 supports CAVLC, and each of the individual components of a given coefficient block is processed in a different processing procedure (coefficient component determination order) from that of the second embodiment. This is different from the second embodiment in that the coefficient level is selected. Since the other configuration is the same as that of the second embodiment, detailed description thereof will be omitted. In the following, the fourth embodiment will be described focusing on differences from the second embodiment.

  FIG. 8 is a flowchart illustrating processing of the quantization unit 13 according to the fourth embodiment.

First, the quantizing unit 13 obtains a scaled coefficient c _i obtained by scaling the input transform coefficient according to the quantization parameter QP (S601). However, i = 0,..., M are numbered in the order of scanning such as Zig-Zag scanning, as in the first and third embodiments, unlike the second embodiment. That is, in the fourth embodiment, sorting based on the magnitude of the absolute value is not performed.

Then, the quantization unit 13, inverse order i = M scan, ..., 0, candidate positions _{i 1} of the last non-zero coefficient is searched, it is set (S602). That is, i ₁ searches for i ₁ that is the maximum value of i for which a _i > T ₀ . Here, _{T 0} is the threshold for selecting the coefficient level values _0, T 0 = 0.5, or _may be such as T 0 = 0.5 + μ. Moreover, it is good also as controllable from the outside. If there is no such i ₁ , the target block may be determined to have no coefficient, and the process may be terminated without performing subsequent processing.

Next, the quantization unit 13 selects the coefficient level li for coefficients that do not include 0 as coefficient level candidates in the reverse scan order i = i ₁ ,..., 0 (S603, S604). . That is, for the absolute value a _i of c _i , t _i is obtained by the following equation (67) as a candidate of the absolute value u _i of the coefficient level l _i , and further, the distortion change amount ΔE by the following equation (68): _i is determined.

t _i = floor (a _i ) (67)
ΔE _i = ΔE (a _i , t _i ) = 1 / 2− (a _i −t _i ) (68)
Here, for _i where ΔE _i ≧ 0, u _i = t _i, and in the subsequent processing, t _i +1 may not be a candidate (however, even if ΔE _i ≧ 0) , T _i = 0, it may be unselected and processed at the time of selecting an unselected coefficient level later).

On the other hand, for _{i where} ΔE _i <0, the amount of change ΔR _i in the code amount when t _i is selected as u _i and when t _i +1 is selected, and RD in the following equation (69) is obtained. The coefficient level is selected from t _i or t _i +1 by judging the sign of the amount of change in cost.

ΔJ _i = ΔE _i + μΔR _i (69)
That is, t _i +1 is selected when ΔJ _i <0. The coefficient level li is successively updated using the selected u _i .

However, when t _i = 0, the selection is not performed, and processing is performed when a subsequent unselected coefficient level is selected. Here, initialization is performed once as in the following expression (70). Here, the initial value of the unselected coefficient level l _i is, for example, f, where f = 1/3 for an intra picture and f = 1/6 for an inter picture. It is also possible to use a method of making it.

u _i = floor (a _i + f) (70)
On the other hand, in the case of t _i > 0, if the variable suffixLength is determined, ΔR _i can be efficiently obtained as a code length difference in the variable length code table selected by the suffix Length.

Here, the △ _{R i} is the difference code lengths only relates coefficient levels _{l i,} as in the second embodiment, the "if there is no change in level_prefix, a △ _R i = 0" Therefore, in this case, the selection can be made only by the determination of ΔE _i <0.

  Here, the variable suffixLength is a variable for selecting a variable-length code table when the coefficient level is CAVLC encoded, and changes in the reverse order of scanning according to the magnitude of the absolute value of the coefficient level. Therefore, by selecting coefficient levels in the reverse order of scanning, suffixLength is also determined. For the unselected coefficient level, the suffixLength is updated according to the initial value. Note that the coefficient level when the accuracy of the suffixLength is low, such as when the number of trailingOnes is less than 3, may be set as unselected and processed when selecting an unselected coefficient level later.

When the processes in steps S603 and 604 described above are completed, the quantization unit 13 performs reverse scan order i = i ₁ , for coefficients that remain unselected in the above process (the processes in steps S603 and 604 described above). .., 0, the coefficient level l _i is selected (S605, S606). In the processing so far, the processing of the selected coefficient component is skipped.

That is, reverse i _{= i} 1 scan, ..., 0, for each i, determine the code amount of the change amount △ _{R i} in the case of selecting a _t i +1 If you choose _{t i} as _{u i,} the following The coefficient level is selected from t _i or t _i +1 by determining the sign of the amount of change in the RD cost in equation (71). That is, t _i +1 is selected when ΔJ _i <0. The coefficient level l _i is sequentially updated using the selected u _i .

ΔJ _i = ΔE _i + μΔR _i (71)
When calculating the change amount ΔR _i of the code amount, it is calculated with reference to the updated coefficient level l _i as it is, if necessary, or if it is updated by the processing so far. △ calculation of R _i, similar to that described in the second embodiment, as selecting t _i, between when selecting t i _+1, only syntax element with a change in the amount of codes may be considered, it is possible to narrow the syntax elements associated with the calculation processing efficiently △ R _i.

(D-2) Effects of Fourth Embodiment According to the fourth embodiment, the following effects can be achieved in addition to the effects of the second embodiment.

  The coefficient sorting process, which was necessary in the second embodiment, is no longer necessary. Further, when setting the initial value of the coefficient level, the coefficient that can be selected only by the code length difference of the coefficient level according to suffixLength is selected. Therefore, the number of coefficients to be selected can be reduced while observing the change in the code amount of the entire block, and the quantization process can be performed with a smaller amount of computation and processing than in the second embodiment.

(E) Fifth Embodiment Hereinafter, a fifth embodiment of a quantization device, a program and a method, and a moving image coding device according to the present invention will be described in detail with reference to the drawings. Note that the quantization apparatus according to the fifth embodiment will be described as a quantization unit.

(E-1) Configuration and Operation of Fifth Embodiment In the fifth embodiment, after a quantization process in units of coefficient blocks as in the first to fourth embodiments, a plurality of coefficient blocks are included. Optimize the quantization result in a larger unit.

  For example, H.M. In the case of H.264 4 × 4 transform coefficients, the presence / absence of coefficients in the unit of 8 × 8 obtained by collecting four 4 × 4 blocks by coded_block_pattern (in the case of none, all coefficient levels of the target block are 0) The amount of rate distortion cost change due to this control is evaluated, and the quantization optimization process is performed.

  In the fifth embodiment, since the quantization unit 13 in the first embodiment is merely replaced with the quantization unit 13A, the description of the configuration other than the quantization unit 13A is omitted.

  FIG. 9 is a block diagram illustrating a functional configuration inside the quantization unit 13A of the fifth embodiment.

  The quantization unit 13A includes a quantization processing unit 135, a distortion change amount total calculation unit 136, a rate change amount calculation unit 137, and a coefficient level selection unit 138.

The quantization processing unit 135 receives transform coefficients in units of blocks, and performs quantization processing by selection of RD optimum coefficient levels in units of blocks as in the first to fourth embodiments according to a predetermined quantization parameter QP. And ΔE _i is output for the obtained coefficient level of the block unit and the coefficient level whose absolute value is 1.

  Hereinafter, as an example, an example of simultaneously optimizing the coefficient levels of four 4 × 4 blocks included in a certain 8 × 8 block expressed by coded_block_pattern will be described. However, the present invention is not limited to this example. The other control methods that can control a plurality of coefficient blocks at the same time without coefficients can be similarly configured.

  The following processing is performed when the coefficient level of the 4 × 4 block included in the 8 × 8 block that can be controlled by coded_block_pattern does not have a coefficient level value larger than the absolute value 1, that is, the coefficient level value is −1, 0. , 1 only. Also, when all coefficient level values are 0, no processing is performed.

The distortion change amount total calculation unit 136 obtains the distortion change amount ΔE by adding all the distortion change amounts ΔE _i related to the coefficient component whose absolute value of the coefficient level value is 1.

That is, obtaining the e _j is the sum of the distortion change amount △ E _i nonzero coefficient related coefficient blocks j by the following (72) below. The time for the i coefficient level value is 0 △ E _i is not added.

e _j = Σ _{i (li ≠ 0)} ΔE _i (72)
The distortion change amount total calculation unit 136 obtains the total distortion change amount ΔE over the entire target block by the following equation (73). ΔE is the amount of change in the scaled distortion from the case where all the target coefficient blocks have no coefficient to the state of the quantization result in units of blocks (including a coefficient having an absolute value of 1).

ΔE = Σ _j e _j (73)
The rate change amount calculation unit 137 obtains, as r _j, the code amount when the coefficient block j is encoded by an entropy encoder such as CABAC or CAVLC as the change amount of the code amount. Then, the rate change amount calculation unit 137 obtains ΔR by the following equation (74) as a difference when not coding all these codewords by using the sum of these and coded_block_pattern. However, the difference between bits () in the second term of equation (74) is the difference in code amount when the presence / absence of the coefficient of the entire 8 × 8 region of the coded_block_pattern is controlled, and the CABAC and CAVLC to be used. It is a value determined by entropy coding. Further, it may be configured that the difference in the code amount is regarded as very small and ignored (not calculated).

ΔR = Σ _j r _j + (bits (CBP = 1) −bits (CBP = 0)) (74)
Then, the coefficient level selection unit 138 uses the obtained ΔE and ΔR and the scaled Lagrangian multiplier μ to make a plurality of blocks have no coefficient at the same time, and the coefficient level by quantization processing for each block. A change amount ΔJ of the RD cost when used is obtained by the following equation (75). When ΔJ ≧ 0, all the target blocks have no coefficient. If ΔJ <0, or if it includes a coefficient level with an absolute value of 2 or more, which is not subject to processing above, the quantization processing result for each block is output as it is.

ΔJ = ΔE + μΔR (75)
(E-2) Effects of Fifth Embodiment According to the fifth embodiment, the following effects can be achieved.

  Optimization processing of the quantization result in a larger unit including a plurality of blocks can be performed, and coding efficiency can be improved. Further, since the evaluation of the RD cost at this time is performed in the scaled coefficient level region and only the change amount is evaluated, the distortion change amount can be evaluated without multiplication, and the processing is small. Can be realized with load.

(F) Other Embodiments The present invention is not limited to the above-described embodiments, and may include modified embodiments as exemplified below.

(F-1) In each of the above embodiments, the moving picture coding apparatus (quantization apparatus) of the present invention is the H.264 standard. However, the present invention is not limited to the above-described embodiment, and is applied to quantization processing in various moving image encoding processes including transform coefficient quantization and entropy encoding. Is possible.

  DESCRIPTION OF SYMBOLS 10 ... Moving image encoding device, 11 ... Prediction difference signal generator, 12 ... Conversion part, 13 ... Quantization part, 14 ... Entropy encoding part, 15 ... Dequantization / inverse conversion part, 16 ... Decoded image generation part , 17 ... Frame buffer, 18 ... Prediction unit, 131 ... Scale calculation unit, 132 ... Distortion change amount calculation unit, 133 ... Rate change amount calculation unit, 134 ... Coefficient level selection unit.

Claims (9)

In a moving image encoding process, in a quantization device that quantizes a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image,
Scaling means for performing scaling for quantization on the transform coefficient;
A first coefficient level compensation obtained by converting the absolute value of the scaled conversion coefficient into an integer by rounding off the decimal point, and a coefficient for determining a second coefficient level candidate whose absolute value is 1 larger than the first coefficient level candidate Level candidate determination means,
When the first coefficient level compensation is selected and when the second coefficient level candidate is selected, the scaled distortion change amount corresponding to the difference in quantization error and entropy encoding are performed. The amount of change in the amount of code required for the above is obtained, and when the first coefficient level compensation is applied using the obtained amount of change in distortion and the amount of change in the amount of code, the second coefficient level 係数Rate distortion cost change amount calculating means for calculating a change amount of the rate distortion cost when the complement is applied; and
A coefficient that selects either the first coefficient level candidate or the second coefficient level candidate as a coefficient level to be applied to quantization of the transform coefficient according to the calculation result of the rate distortion cost change amount calculation means. And a level selection means. The quantization apparatus entropy-encodes the quantized transform coefficient in the moving image encoding process, and represents that the coefficient is a final non-zero coefficient in the scan order of the transform coefficient. Supplying the quantized transform coefficient to the encoding means having the syntax;
The scaling means scales the transform coefficient according to a quantization parameter,
The rate distortion cost change amount calculation means determines a range for searching for a final non-zero coefficient component position in each coefficient component of the conversion coefficient, and a certain final non-zero coefficient component position in the forward direction of the scan with respect to this range. And the amount of change in rate distortion cost when another final non-zero coefficient component position is selected, and the final non-zero coefficient component position is determined according to the sign of the rate distortion cost change amount. The rate when the first coefficient level compensation is applied and the second coefficient level compensation is applied to each coefficient component in the reverse direction of scanning from the determined final non-zero coefficient component position. Find the amount of change in distortion cost,
The coefficient level selection means selects a coefficient level to be applied to each coefficient component of the transform coefficient according to the sign of the rate distortion cost change amount calculated by the rate distortion cost change amount calculation means. The quantization apparatus according to claim 1. The quantization apparatus entropy-encodes the quantized transform coefficient in the moving image encoding process, and the quantized transform is performed to an encoding unit including run-length encoding of a zero coefficient. Supply coefficient,
The scaling means scales the transform coefficient according to a quantization parameter,
The rate distortion cost change amount calculating means sorts the coefficient components of the transform coefficients scaled by the scaling means in ascending order of absolute values, and selects the first coefficient level candidates and the second coefficient level candidates. , When the first coefficient level compensation is applied to each coefficient component of the transform coefficient in the sorted order for each coefficient component of the transform coefficient, and the second coefficient level. The amount of change in rate distortion cost with respect to the case where compensation is applied is obtained, and the coefficient level selection means determines the conversion coefficient according to the sign of the amount of change in rate distortion cost calculated by the rate distortion cost change amount calculation means. 2. The quantization apparatus according to claim 1, wherein a coefficient level to be applied to each coefficient component is selected. The quantization apparatus performs entropy encoding on the quantized transform coefficient, and has encoding syntax indicating that it is a final non-zero coefficient in the scan order of each coefficient component of the transform coefficient To supply the quantized conversion coefficient,
The scaling means scales the transform coefficient according to a quantization parameter,
The rate distortion cost change amount calculating means determines a search start position of the final non-zero coefficient position in the reverse order of scanning of each coefficient component of the transform coefficient, and sets 0 as a coefficient level candidate in the reverse order of the scan from the search start position. Determine the rate distortion cost change amount when selecting a certain final non-zero coefficient component position and another final non-zero coefficient component position up to the coefficient component that is not, and determine the sign of this rate distortion cost change amount The final non-zero coefficient component position to be applied is determined, the coefficient level of the search range is determined, and the first coefficient level compensation is applied to each coefficient component in the reverse direction of the scan from the search stop position. The amount of change in the rate distortion cost between the case and the case where the second coefficient level compensation is applied,
The coefficient level selection means selects a coefficient level to be applied to each coefficient component of the transform coefficient according to the sign of the rate distortion cost change amount calculated by the rate distortion cost change amount calculation means. The quantization apparatus according to claim 1. The quantization device entropy-encodes the transform coefficient quantized by the quantization device, and includes zero-length run-length coding and absolute coefficient level appearing in the order of quantization coefficient level coding. Depending on the magnitude of the value, the quantized transform coefficient is supplied to an encoding means including encoding in which the code used for encoding of the non-zero coefficient level is changed,
The scaling means scales the transform coefficient according to a quantization parameter,
The rate distortion cost change amount calculation means is either the first coefficient level candidate or the second coefficient level candidate corresponding to the coefficient component in the coding order of the coefficient level for each coefficient component of the transform coefficient. When the coefficient level is zero, the coefficient level applied to the coefficient component is initialized to a value quantized by a predetermined quantization process, and the code used to encode the coefficient component is determined. For the coefficient component, the coefficient level is selected by obtaining the amount of change in the rate distortion cost when the first coefficient level compensation is applied and when the second coefficient level compensation is applied, Further, for the coefficient component for which the rate distortion cost change amount has not yet been obtained by the above processing, the rate distortion cost change amount is obtained,
The coefficient level selection means selects a coefficient level to be applied to each coefficient component of the transform coefficient according to the sign of the rate distortion cost change amount calculated by the rate distortion cost change amount calculation means. The quantization apparatus according to claim 1. In a moving image encoding process, in a quantization device that quantizes a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image,
The quantization apparatus performs entropy coding on the quantized result and quantizes the coding means including coding capable of expressing that the coefficient levels are all zero over a plurality of transform coefficient blocks. Supply the above conversion factor,
Unit quantization means for quantizing by the quantizing device according to any one of claims 1 to 5 for each transform coefficient block constituting a plurality of transform coefficient blocks to be quantized,
A distortion change amount sum calculation for obtaining a sum of scaled strain change amounts from the case where coefficient level 0 is applied to a transform coefficient block quantized by the unit quantization means and having a coefficient level of non-zero. Means,
A code amount required for entropy encoding the transform coefficient block quantized by the unit quantization means, and a code amount for expressing that the coefficient levels of the transform coefficient blocks are all zero. A code amount change amount calculating means for obtaining a change amount of
The transform quantized by the unit quantization means using the total distortion change amount obtained by the distortion change amount summation calculating means and the code amount change amount obtained by the code amount change amount calculating means. The amount of change in the rate distortion cost between the coefficient block and the case where the coefficient levels of each of the transform coefficient blocks are all zero is obtained, and the quantization result of each of the transform coefficient blocks is determined according to the amount of change in the rate distortion cost. , Coefficient level selection means for determining whether the coefficient levels are all zero or the quantization result of the unit quantization means;
A quantization apparatus comprising: In a moving image encoding process, in a quantization program for quantizing a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image,
Computer
Scaling means for performing scaling for quantization on the transform coefficient;
A first coefficient level compensation obtained by converting the absolute value of the scaled conversion coefficient into an integer by rounding off the decimal point, and a coefficient for determining a second coefficient level candidate whose absolute value is 1 larger than the first coefficient level candidate Level candidate determination means,
When the first coefficient level compensation is selected and when the second coefficient level candidate is selected, the scaled distortion change amount corresponding to the difference in quantization error and entropy encoding are performed. The amount of change in the amount of code required for the above is obtained, and when the first coefficient level compensation is applied using the obtained amount of change in distortion and the amount of change in the amount of code, the second coefficient level 係数Rate distortion cost change amount calculating means for calculating a change amount of the rate distortion cost when the complement is applied; and
A coefficient that selects either the first coefficient level candidate or the second coefficient level candidate as a coefficient level to be applied to quantization of the transform coefficient according to the calculation result of the rate distortion cost change amount calculation means. A quantization program that functions as a level selection means. In a moving image encoding process, in a quantization method for quantizing a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image,
A scaling step for scaling the transform coefficient for quantization;
The first coefficient level compensation obtained by converting the absolute value of the scaled conversion coefficient into an integer by rounding down the decimal point, and the second coefficient level candidate having a larger absolute value than the first coefficient level candidate by 1 A coefficient level candidate determination step;
When the first coefficient level compensation is selected and when the second coefficient level candidate is selected, the scaled distortion change amount corresponding to the difference in quantization error and entropy encoding are performed. The amount of change in the amount of code required for the above is obtained, and when the first coefficient level compensation is applied using the obtained amount of change in distortion and the amount of change in the amount of code, the second coefficient level 係数A rate distortion cost change amount calculating step for calculating a change amount of the rate distortion cost when the supplement is applied; and
A coefficient for selecting either the first coefficient level candidate or the second coefficient level candidate as a coefficient level to be applied to quantization of the transform coefficient according to the calculation result of the rate distortion cost change amount calculating step. And a level selection step.   In the moving image coding apparatus, the moving image coding apparatus includes a quantization device that quantizes a transform coefficient related to an error signal between an input image constituting a moving image to be encoded and a predicted image of the input image, and the quantization device includes: A moving picture coding apparatus to which the quantization apparatus according to any one of claims 1 to 6 is applied.

Images