JP4709074B2 - Moving picture encoding method, apparatus, program thereof, and recording medium recording the program - Google Patents

Description

  The present invention relates to a high-efficiency image signal encoding technique, and more particularly, to a moving image encoding method considering visual sensitivity.

  H. In H.264, with the introduction of intra prediction and variable shape motion compensation, the types of prediction modes are increasing as compared with the conventional standardized method. For this reason, in order to reduce the amount of codes while maintaining a constant subjective image quality, it is necessary to select an appropriate prediction mode. H. In the H.264 reference software JM [Non-Patent Document 1], the following prediction mode that minimizes the RD cost is selected.

Here, S is an original signal, q is a quantization parameter, m is a number indicating a prediction mode, and ^ Sm, q is a prediction for mode S using mode m and is quantized using q. It is a decoded signal. Λ is a Lagrange's undetermined multiplier used for mode selection. Further, D (S ,， Sm, q) is a sum of square errors shown in the following equation.

Here, S ^Y , S ^U , and S ^V are Y, U, and V components of the original signal, and ^ Sm, q ^Y , ^ Sm, q ^U , ^ Sm, q ^V are Y, U, and V of the decoded signal. It is an ingredient.

  H. The calculation of the decoded signal in H.264 is shown below. The symbols used for explanation are summarized in Table 1.

H. In the H.264 encoding process, orthogonal transformation using the transformation matrix Φ is performed on the prediction error signal R (= S−P _m ) when the prediction of the mode number m is used as follows.

C = ΦRΦ ^t (1)
Φ ^t represents a transposed matrix with respect to the transformation matrix Φ. The transformation matrix Φ is an orthogonal matrix of integer elements expressed by the following equation.

Next, since the matrix Φ is a non-normal matrix, processing equivalent to matrix normalization is performed.

C _n = N (C) (3)
Further, the quantization using the quantization parameter q is performed on C as follows. In JM, normalization is built into quantization.

V = Q (C _n ) (4)
On the other hand, H. In the H.264 decoding process, V is inversely quantized as in the following equation to obtain a decoded value of the transform coefficient.

Next, the inverse transformation is applied to ^ C _q as shown in the following equation to obtain a prediction error decoded signal.

Finally, a decoded signal of the encoding target image is obtained by the following equation.

In addition to the above non-patent document 1, there are the following non-patent documents 2 to 5 as reference documents describing the technology related to the present invention. Further, in Patent Document 1 below, in order to improve the encoding efficiency by a plurality of encodings without a significant increase in hardware, the motion amount of a moving image is detected from the motion vector detected during the first encoding. Is described, and the coding rate at the time of the second coding is controlled in consideration of human visual characteristics and coding distortion reduction.
JP-A-10-336675 KPLim and G. Sullivan and T. Wiegand, Text Description of Joint Model Reference Encoding Methods and Decoding Concealment Methods. Joint Video Team (JVT) of ISO / IEC MPEG and ITU-T VCEG, JVT-R95, Jan., 2006. JLMannos and DJSakrison. The effect of a visual fidelity criterion on the encoding of images. IEEE Trans. Infomation Theory, Vol.IT-20, pp.523-536, July 1974. NBNill. A visual model weighted cosine transform for image compression and quality assessment. IEEE Trans. Commun., Vol.COM-33, No.12, pp.551-557, June 1985. B. Chitprasert and KRRao. Human visual weighted progressive image transmission. IEEE Trans. Commun., Pp.1040-1044, July 1990. Hitoshi Uekura, Hiroshi Watanabe, Naoki Kobayashi, Susumu Ichinose, Hiroshi Yasuda. Global motion / brightness change compensation video coding considering computational complexity. IEICE Transactions, Vol. .1676-1688, Spt. 1999.

  As described above, the subjective image quality scale used in JM is a square error. However, this square error is not necessarily a distortion amount reflecting subjective image quality degradation. For example, a change in a high frequency component is less visually detected than a change in a low frequency component. It is also known that image quality degradation in a fast-moving scene is relatively less noticeable than image quality degradation in a still scene (time masking effect). However, an encoder (for example, JM) that does not use such visual characteristics still has room for improvement in terms of efficient code amount reduction.

  In the technique described in Patent Document 1, encoding rate control such as quantization step width control in consideration of visual characteristics is performed, but it does not support selection of an encoding mode.

  The present invention has been made in view of such circumstances, and a prediction mode selection method for efficiently reducing a code amount by using a distortion amount reflecting subjective image quality in consideration of a masking effect in time and space. The purpose is to establish.

  The above-mentioned spatial masking effect is derived from the difference in visual sensitivity with respect to the spatial frequency component. Therefore, the distortion amount corresponding to the subjective image quality is defined by weighting the orthogonal transformation coefficient according to the visual sensitivity for each frequency component. Further, in consideration of time masking, the above-described weighted distortion amount is further weighted according to the motion amount. The amount of distortion weighted according to such spatiotemporal properties is used within the RD cost.

That is, the present invention sets a prediction mode such as a block size for intra-frame prediction and inter-frame prediction, and performs video compression that performs information compression by orthogonal transformation and quantization on a prediction error signal obtained by the prediction. When determining the prediction mode based on the Lagrangian cost function consisting of distortion, code amount, and undetermined multiplier , the visual component calculated by a predetermined visual sensitivity function is applied to the frequency components in the block subjected to orthogonal transformation. The process of calculating the ratio between the product of the sensitivity value and the frequency component in the block subjected to the orthogonal transformation as a sensitivity coefficient, and weighting obtained by multiplying the square error by the sensitivity coefficient for each prediction mode a process of calculating the distortion amount, setting the cost function using the weighted distortion amount, the process of calculating the cost of each prediction mode by using the cost function Characterized by having a step of selecting a prediction mode that minimizes the cost calculated by using the cost function.

In the above present invention, the visual sensitivity function has at least the parameter distance parameter view obtained by dividing the distance viewing as one by the width of the image, is a decreasing function of the visual distance parameter, when calculating the sensitivity coefficient , depending on the global motion of the frame, visual distance parameter when the motion amount is large, and set to a value larger than the viewing distance parameter when the motion amount is small, the sensitivity coefficients for each frame It is characterized by being adaptively changed.

In the above present invention, the visual sensitivity function has at least the parameter distance parameter view obtained by dividing the distance viewing as one by the width of the image, is a decreasing function of the visual distance parameter, when calculating the sensitivity coefficient , according to the motion amount in the block, visual distance parameter when the motion amount is large, and set to a value larger than the viewing distance parameter when the motion amount is small, adaptive sensitivity coefficients for each block It is characterized by changing to.

  In the present invention, the RD cost of mode selection is adaptively switched in consideration of the degree of contribution of coding distortion to subjective image quality. As a result, weighting that emphasizes code amount reduction is added to the RD cost for an area where the contribution of coding distortion to the subjective image quality is small. In this case, since the code amount is reduced for a region that is difficult to detect visually, the code amount can be efficiently reduced while maintaining the subjective image quality of the decoded image.

  Hereinafter, specific embodiments of the present invention will be described in detail.

[Quantization error signal weighting]
First, the weighting based on visual sensitivity for the quantization error signal will be described. In the present embodiment, the following RD cost is used.

The following weighted distortion amount is used as the distortion amount used for the calculation of the RD cost.

Here, C _n ^{Y (i)} [k, l], C _n ^{U (i)} [k, l], C _n ^{V (i)} [k, l] are macroblocks (16 × 16 in the case of Y component ⁾ . In the case of [pixel], U, and V components, it is the i-th sub-block scanned in raster scanning among the sub-blocks (N × N [pixel]) in 8 × 8 [pixel]). ^ C _q ^{Y (i)} [k, l], ^ C _q ^{U (i)} [k, l], ^ C _q ^{V (i)} [k, l] are decoding transform coefficients (Y Among the sub-blocks (N × N [pixels]) within the 16 × 16 [pixel] in the case of the component and 8 × 8 [pixel] in the case of the U and V components, the i th sub-scan in the raster scan is performed. It is a block. Further, W _{k, l} ^Y , W _{k, l} ^U , W _{k, l} ^V are weighting coefficients set to 1 or less, and are hereinafter referred to as sensitivity coefficients. The calculation of the sensitivity coefficient will be described in detail in [Calculation of sensitivity coefficient].

In the above equation, setting W _{k, l} ^Y , W _{k, l} ^U , W _{k, l} ^V to a small value is equivalent to estimating the quantization distortion D (C _n , _Cq ) small. From the normality of orthogonal transformation, if W _{k, l} ^Y = 1, W _{k, l} ^U = 1, W _{k, l} ^V = 1 for all _{k, l} , the weighted distortion described above is obtained. The quantity is equivalent to the sum of squared errors.

[Calculation of sensitivity coefficient]
If the k-th column vector (N-dimensional vector) of the transformation matrix Φ (N × N matrix) is φ _k , a base image for the matrix can be obtained from the following equation. H. In the case of H.264, possible values for N are either 4 or 8.

Here, φ _l ^t is a transposed vector of φ _l .

Each base image f _{k, l} (x, y) (0 ≦ x, y ≦ N−1) is subjected to a discrete Fourier transform represented by the following equation to obtain a Fourier coefficient. In the following, N = 2 ^m .

Here, j is an imaginary unit. The following weighting is performed on the obtained Fourier coefficient F _{k, l} (u, v) (0 ≦ u ≦ N−1, 0 ≦ v ≦ N−1).

Hereinafter, ˜F _{k, l} (u, v) (where ˜ is a spelling symbol attached to the upper side of F) will be described. ^ G (η) is a function represented by the following equation using g (η).

Here, α is a parameter for controlling the weight and is given from the outside. Further, g (η) is a function known as a visual sensitivity function, and is expressed in a function form as shown in the following equation.

g (η) = (a + bη) exp (− (cη) ^d ) (12)
Here, a, b, c, and d are parameters (hereinafter referred to as model parameters) that define the function form of the visual sensitivity function, and take the following values, for example. The following values are based on Non-Patent Literature 2, Non-Patent Literature 3, and Non-Patent Literature 4 in order from the top.

(A, b, c, d) = (0.4992, 0.2964, −0.114, 1.1) (13)
(A, b, c, d) = (0.2, 0.45, −0.18, 1) (14)
(A, b, c, d) = (0.246, 0.615, -0.25, 1) (15)
In Equation (12), η is the following value.

η ₀ is an argument for which g (η) takes a maximum value as follows.

θ (r, H) is an angle per pixel when an image having a vertical width H is observed at a viewing distance rH, and is given by the following equation.

Hereinafter, r is referred to as a viewing distance parameter. From the above discussion, it can be seen that ˜F _k (u, v) is a function of the viewing distance parameter r. Therefore, in the following, a notation that explicitly indicates that it is a function of the viewing distance parameter r is used as ˜F _k (u, v, r).

A sensitivity coefficient for the base image f _{k, l} (x, y) (0 ≦ x, y ≦ N−1) is defined as a power ratio of the following equation.

The sensitivity coefficient can be obtained independently of the encoding target image. Therefore, if the sensitivity coefficient is obtained in advance before encoding and stored in the lookup table, the calculation for calculating the sensitivity coefficient at the time of encoding can be omitted. Note that W _{k, l} ^U (r) and W _{k, l} ^V (r) can be similarly obtained. At this time, it is also possible to change the model parameter with the luminance component and the color difference component.

At this time, W _{k, l} ^Y (r), W _{k, l} ^U (r), and W _{k, l} ^V (r) satisfy the following expression.

W _{k, l} ^Y (r) ≦ 1
W _{k, l} ^U (r) ≦ 1
W _{k, l} ^V (r) ≦ 1
this is,
g (η) ≦ 1
Because

It is because it becomes. Since ~ F _k (u, v, r) is a decreasing function of r, W _{k, l} ^Y (r), W _{k, l} ^U (r), W _{k, l} ^V (r) It can be seen that this is a decreasing function for the distance parameter r. This corresponds to the visual characteristic that the visual sensitivity decreases with the viewing distance.

[Correction considering the effect of time masking effect]
When a large change in the time axis direction (high-speed camera pan / tilt, scene change, etc.) occurs in a certain frame, the sensitivity to the image quality of that frame extremely decreases. This is a visual characteristic known as the time masking effect. Therefore, the sensitivity coefficient is controlled to be a small value for a frame in which a large change in the time axis direction has occurred. In the present embodiment, a viewing distance parameter is used for controlling the sensitivity coefficient.

For example:
(1) When controlling the viewing distance parameter for each frame:
r = r ₁ ... When the magnitude of the global motion vector is greater than or equal to the threshold value r = r ₂ ... Otherwise, (20)
Here, r ₁ > r ₂ . Note that the global motion vector is calculated by the method of Non-Patent Document 5, for example.
(2) When controlling the viewing distance parameter for each macroblock:
r = r ₁ ′ ... when the size of the motion vector of the macroblock is greater than or equal to the threshold value r = r ₂ ′ ... otherwise (21)
Note that when sub-macroblocks in a macroblock have different motion vectors, the following occurs.

r = r ₁ ′… when the minimum value of the magnitude of the motion vector included in the sub macroblock is equal to or greater than the threshold value r = r ₂ ′ —otherwise (22)
Here, r ₁ '> r ₂ '.

[Processing flowchart]
The processing of the embodiment of the present invention will be described with reference to the processing flowchart of FIG. Since processes other than the selection of the prediction mode are the same as those of the conventional moving image encoding process, the description of the processing parts other than the selection of the prediction mode is omitted.

  Step S1: Write the initial value of the prediction mode to the register X.

  Step S2: A register C for storing the minimum cost and a register M for storing the optimum mode are respectively set to initial values.

  Step S3: The prediction mode, the quantization parameter, the encoding target signal, and the reference signal are input, the code amount when the given prediction mode is used is calculated, and the calculated value is written to the register. The specific calculation method follows the JM method.

  Step S4: The prediction mode, the quantization parameter, the encoding target signal, and the reference signal are input, the weighted distortion amount when the given prediction mode is used is calculated, and the calculated value is written to the register. Details of this processing will be described later with reference to FIG.

  Step S5: The prediction mode and the quantization parameter are input, the undetermined multiplier is calculated, and the calculated value is written to the register. The specific calculation method follows the JM method.

  Step S6: The code amount, the weighted distortion amount, and the undetermined multiplier are input, the RD cost is calculated, and the calculated value is written to the register. The specific calculation method follows equation (8).

  Step S7: Using the RD cost calculated in step S6 and the value of the register C as inputs, it is determined whether or not the RD cost calculated in step S6 is smaller than the value of the register C. A certain boolean value is output. If the output is a true value, the process proceeds to step S8. If the output is a false value, the process proceeds to step S10.

  Step S8: The RD cost calculated in step S6 is written in the register C.

  Step S9: Write the value of the register X to the register M.

  Step S10: It is determined whether or not the above processing has been completed for all prediction modes, and if not, the process proceeds to step S11. When the process is completed for all prediction modes, the process is terminated. The mode finally stored in the register M is the prediction mode selected as the optimum mode, and the cost at that time is the value stored in the register C.

  Step S11: A value representing the next prediction mode is written in the register X. The order of prediction modes to be written to the register is given in advance. Then, it returns to step S3 and repeats a process similarly.

  A detailed processing flow of step S4 in FIG. 1 is shown in FIG.

  Step S21: The conversion coefficient obtained by the conversion matrix Φ is input, normalization processing is performed, and the normalized conversion coefficient is written in a register as a one-dimensional array. It should be noted that a scanning method in which a conversion coefficient that is two-dimensional data is used as one-dimensional data is given from the outside. The same applies to the following steps S22 and S24.

  Step S22: The quantization result for the transform coefficient is input, inverse quantization processing is performed, and the decoded value of the transform coefficient is written to the register as a one-dimensional array.

  Step S23: A motion vector is input, a process for calculating a viewing distance parameter is performed according to the motion vector, and the calculated viewing distance parameter is written to a register. For example, the method shown in equations (20) to (22) is followed. Here, an example is shown in which two types of sensitivity coefficients are selected, but the present invention can be similarly applied even when there are three or more selection candidates. Note that the motion vector used as an input here is given separately. For example, there is a method using a motion vector calculated in the process of encoding (using an average vector of motion vectors in a macroblock). Alternatively, there is a method of using a motion vector for a macroblock (16 × 16 [pixel]) obtained independently of encoding.

  Step S24: A processing for calculating a sensitivity coefficient according to the viewing distance parameter is performed with the viewing distance parameter as an input, and the calculated sensitivity coefficient is written in a register as a one-dimensional array. The specific calculation method follows equation (19). If selectable viewing distance parameters are determined in advance, the sensitivity coefficient can be calculated in advance, so that the sensitivity coefficient calculation process can be omitted. In this case, this processing is as follows. With the viewing distance parameter as an input, the sensitivity coefficient previously stored in the register is read in accordance with the viewing distance parameter, and is written in the register as a one-dimensional array.

  Step S25: The value of the counter i used for counting the number of repetitions is initialized to zero. The value of the register S is initialized to 0.

  Step S26: The i-th component of the normalized transform coefficient output in step S21 and the i-th component of the decoded value of the transform coefficient output in step S22 are input, the difference between them is obtained, and the difference value is obtained. A process of squaring is performed, and the calculated square error (coding distortion of the transform coefficient) is written to the register.

  Step S27: The square error output at step S26 and the i-th component of the sensitivity coefficient are input, the square error is multiplied by the sensitivity coefficient, and the multiplication result is written to the register.

  Step S28: The value calculated in step S27 and the value of the register S are input, both are added, and the addition result is written to the register S.

  Steps S29 and S30: The above processing is repeated while adding 1 to the counter i, and is performed for all components of the transform coefficient.

[Device configuration example]
An apparatus configuration diagram according to the embodiment of the present invention is shown in FIG. Note that the configuration of the processing unit that encodes a moving image using the optimal prediction mode selected according to the present embodiment is the same as the configuration of a conventional general moving image encoding apparatus, and therefore here, the prediction is performed. Only the part of the processing configuration for selecting the mode will be described.

  Initial mode setting unit 101: Writes the initial value of the prediction mode to the mode storage unit 102.

  Code amount calculation unit 103: Inputs a prediction mode, a quantization parameter, an encoding target signal, and a reference signal, calculates a code amount when encoding, and writes the calculated value to the code amount storage unit 104. The specific calculation method follows the JM method.

  Weighted distortion amount calculation unit 105: The prediction mode, the quantization parameter, the encoding target signal, and the reference signal are input, the weighted distortion amount when encoding is calculated, and the calculated value is used as the weighted distortion amount storage unit 106. Export to Details of this processing will be described later with reference to FIG.

  Undetermined multiplier calculation unit 107: Inputs the prediction mode and the quantization parameter, calculates the undefined multiplier, and writes the calculated value to the undefined multiplier storage unit 108. The specific calculation method follows the JM method.

  Cost calculation unit 109: The code amount storage unit 104, the weighted distortion amount storage unit 106, the code amount read from the undetermined multiplier storage unit 108, the weighted distortion amount, and the undetermined multiplier are input to calculate the RD cost. The calculated value is written in the cost storage unit 110. The specific calculation method follows equation (8).

  Minimum cost determination unit 111: The RD cost and the minimum cost read from the cost storage unit 110 and the minimum cost storage unit 112 are input, and it is determined whether the RD cost is smaller than the minimum cost. Outputs a boolean value that is When the output is a true value, the RD cost read from the cost storage unit 110 is written to the minimum cost storage unit 112 and the process proceeds to the optimum mode update unit 113. If the output is a false value, the process proceeds to the final mode determination unit 115.

  Optimal mode update unit 113: Writes the mode used for RD cost calculation written in the minimum cost storage unit 112 into the optimal mode storage unit 114.

  Final mode determination unit 115, mode setting unit 116: The above processing is performed for all prediction modes.

  Optimal mode output unit 117: When processing for all prediction modes is completed, the prediction mode stored in the optimal mode storage unit 114 is output as the optimal mode.

  FIG. 4 shows a detailed configuration of the weighted distortion amount calculation unit 105 in FIG.

  Conversion coefficient normalization unit 201: The conversion coefficient obtained from the conversion matrix Φ is input, normalization processing is performed, and the normalized conversion coefficient is written to the normalized conversion coefficient storage unit 202.

  Transform coefficient decoding unit 203: The quantization result for the transform coefficient is input, inverse quantization processing is performed, and the decoded value of the transform coefficient is written in the decoded transform coefficient storage unit 204. The quantization here includes normalization processing as in the case of JM.

  Distortion amount calculation unit 205: The normalized conversion coefficient storage unit 202 and the decoded conversion coefficient storage unit 204 read from the normalized conversion coefficient and the decoded value of the conversion coefficient are input, and the difference between the two is obtained for each component. A process of squaring the difference value is performed, and the calculated square error is written in the distortion amount storage unit 206.

  Sensitivity coefficient multiplication unit 207: The sensitivity coefficient read from one of the sensitivity coefficient storage units 208-1 and 208-2 and the square error read from the distortion amount storage unit 206 are input, and the square error is multiplied by the sensitivity coefficient. The obtained weighted distortion amount is written in the distortion amount storage unit 212. Although an example of selecting two types of sensitivity coefficients is shown here, the present invention can be similarly applied even when there are three or more selection candidates.

  Viewing distance parameter calculation unit 209: With a motion vector as an input, processing for calculating a viewing distance parameter is performed according to the motion vector, and the calculated viewing distance parameter is written in the viewing distance parameter storage unit 210. For example, the method shown in equations (20) to (22) is followed. Although an example in which two types of viewing distance parameters are selected is shown here, the present invention can be similarly applied even if there are three or more selection candidates. The motion vector used as an input here is separately given. For example, there is a method of using an average vector of motion vectors in a macroblock using a motion vector calculated in the encoding process. Alternatively, a motion vector for a macroblock (16 × 16 [pixel]) obtained independently of encoding may be used.

  Sensitivity coefficient selection unit 211: receives a viewing distance parameter, performs a process of selecting a sensitivity coefficient according to the viewing distance parameter, and outputs a control signal indicating that the selected sensitivity coefficient is input to the sensitivity coefficient multiplication unit 207 To do.

  Distortion amount sum calculation unit 213: The weighted distortion amount of each component read from the distortion amount storage unit 212 is input, the total sum for all components is calculated, and the total sum of the calculation results is the weighted distortion amount of FIG. Write to the storage unit 106.

  The above moving image encoding processing can be realized by a computer and a software program, and the program can be provided by being recorded on a computer-readable recording medium or provided through a network. .

  By using the above-described means, the present invention is, for example, H.264. In an encoder having many encoding modes such as H.264, an appropriate encoding mode can be selected, and encoding efficiency can be improved while maintaining the subjective image quality of the decoded image. H. The feature of high-efficiency image signal encoding such as H.264 is the selection of an encoding mode with a high degree of freedom. Therefore, the encoding performance can be maximized only by selecting an appropriate encoding mode.

It is a processing flowchart of an embodiment of the present invention. It is a process flowchart of step S4 in FIG. It is a device block diagram concerning the embodiment of the present invention. FIG. 4 is a detailed configuration diagram of a weighted distortion amount calculation unit in FIG. 3.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Initial mode setting part 102 Mode memory | storage part 103 Code amount calculation part 104 Code amount memory | storage part 105 Weighted distortion amount calculation part 106 Weighted distortion amount memory | storage part 107 Undecided multiplier calculation part 108 Undecided multiplier memory | storage part 109 Cost calculation part 110 Cost memory | storage Unit 111 minimum cost determination unit 112 minimum cost storage unit 113 optimal mode update unit 114 optimal mode storage unit 115 final mode determination unit 116 mode setting unit 117 optimal mode output unit 201 conversion coefficient normalization unit 202 normalization conversion coefficient storage unit 203 conversion Coefficient decoding unit 204 decoding transform coefficient storage unit 205 distortion amount calculation unit 206 distortion amount storage unit 207 sensitivity coefficient multiplication unit 208-1, 208-2 sensitivity coefficient storage unit 209 viewing distance parameter calculation unit 210 viewing distance parameter storage unit 211 sensitivity coefficient Selection unit 212 Distortion amount Section 213 the amount of distortion sum calculation unit

Claims (6)

In a video encoding method for setting a prediction mode for intra-frame prediction and inter-frame prediction, and compressing information by orthogonal transform and quantization for a prediction error signal obtained by prediction in the set prediction mode,
When determining the prediction mode based on the Lagrangian cost function consisting of distortion, code amount and undetermined multiplier,
The ratio of the frequency component in the block subjected to orthogonal transformation multiplied by the visual sensitivity value calculated by a predetermined visual sensitivity function and the frequency component in the block subjected to orthogonal transformation is calculated as a sensitivity coefficient. The process of
For each prediction mode, and the process of calculating the weighted distortion amount obtained by multiplying the sensitivity coefficient to the square error,
Setting a cost function using the weighted distortion amount, and calculating a cost of each prediction mode using the cost function;
And a step of selecting a prediction mode that minimizes the cost calculated using the cost function. The moving image encoding method according to claim 1,
The visual sensitivity function has a visual distance parameter obtained by dividing the visual distance by the width of the image as at least one parameter, and is a decreasing function of the visual distance parameter,
When calculating the sensitivity coefficients, depending on the global motion of the frame, visual distance parameter when the motion amount is large, and set to a value larger than the viewing distance parameter when the motion amount is small The moving picture coding method characterized by adaptively changing the sensitivity coefficient for each frame. The moving image encoding method according to claim 1,
The visual sensitivity function has a visual distance parameter obtained by dividing the visual distance by the width of the image as at least one parameter, and is a decreasing function of the visual distance parameter,
When calculating the sensitivity coefficients, depending on the amount of motion within the block, visual distance parameter when the motion amount is large, and set to a value larger than the viewing distance parameter when the motion amount is small, the sensitivity A moving picture coding method characterized by adaptively changing a coefficient for each block. In a video encoding apparatus that sets a prediction mode for intra-frame prediction and inter-frame prediction, and performs information compression by orthogonal transform and quantization on a prediction error signal obtained by prediction in the set prediction mode,
Means for calculating the amount of code for each prediction mode;
The ratio of the frequency component in the block subjected to orthogonal transformation multiplied by the visual sensitivity value calculated by a predetermined visual sensitivity function and the frequency component in the block subjected to orthogonal transformation is calculated as a sensitivity coefficient. and, means for calculating for each prediction mode, the weighted distortion amount obtained by multiplying the sensitivity coefficient to the square error,
For each prediction mode, a means for calculating the Lagrange's undetermined multiplier with the quantization parameter as input,
Means for calculating the cost of each prediction mode using a cost function determined by the code amount, the weighted distortion amount, and the undetermined multiplier;
Means for selecting a prediction mode that minimizes the cost calculated using the cost function. A moving picture coding apparatus, comprising: A moving picture coding program for causing a computer to execute the moving picture coding method according to any one of claims 1 to 3 . A computer-readable recording medium recording a moving image encoding program for causing a computer to execute the moving image encoding method according to any one of claims 1 to 3 .

Images