JPH04219074A - Picture coder - Google Patents

Abstract

PURPOSE:To offer a picture coder in which a highest coding efficiency is obtained regardless of a low bit rate with a simple hardware. CONSTITUTION:The picture coder is provided with a prediction device 7 predicting stepwise a 1st picture element going to be coded in a noted block of an input picture signal by using a 2nd picture element already coded to obtain a predicted value, a subtractor 4 obtaining a prediction error of the predicted value obtained from the prediction device 7, a quantizer 5 quantizing the prediction error, a coder 10 coding a quantized value obtained by the quantizer 5, and a correction circuit 8 correcting a 2nd picture element used by the prediction device 7 when the quantized value is a prescribed value or over.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image encoding device for efficiently encoding image signals.

0002

2. Description of the Related Art Image encoding technology has been used for image transmission, recording, and the like. Image encoding is performed for the purpose of transmitting still images as fast as possible in cases such as facsimiles, and for video transmissions such as video conferencing, the purpose is to transmit images using as narrow a band or as low a bit rate as possible. It is done as a purpose. Furthermore, when recording image information on a disk, memory, etc., image encoding technology is used to efficiently record as many images as possible.

[0003] Techniques such as predictive coding, transform coding, and vector quantization are known for image coding. Predictive coding is a method in which a pixel value to be currently coded is predicted from surrounding pixel values that have already been coded, and the prediction error is quantized and coded. Transform coding is a method in which an image is subjected to orthogonal transformation such as Fourier transform or cosine transform to cause energy bias, and the energy bias is used to encode only regions with large energy. Vector quantization is a method in which patterns that often appear in an image are registered in a codebook in advance, the pattern closest to the pattern in the image to be encoded is searched from the codebook, and that pattern number is encoded. be.

Among these methods, predictive coding is not sufficient in terms of coding efficiency to enable transmission of high quality images even at low transmission rates. Transform coding and vector quantization methods have the advantage of higher coding efficiency than predictive coding, but they require a very large amount of calculation and require large hardware. When incorporating an encoding device into a video conference system or the like, it is important to have high encoding efficiency, but miniaturization and cost reduction are also important issues. Furthermore, conventional transform coding has the disadvantage that when the transmission rate becomes low, the subjective quality deteriorates significantly due to block distortion. Vector quantization has the disadvantage that when encoding is performed directly at the level of the image signal, it becomes difficult to represent various images with limited image patterns within the codebook, resulting in poor encoding efficiency.

[0005] Furthermore, a method is known in which the efficiency of vector quantization is improved by vector quantizing a prediction residual signal by combining predictive coding and vector quantization. However, this method is also unable to fully exploit the performance of vector quantization due to the low coding efficiency of conventional predictive coding.

Furthermore, as a predictive coding method, an extrapolation-interpolation predictive coding method that requires very little calculation amount and has high coding efficiency (IEICE Spring National Conference, D-86, 198
9) has been proposed. Although this method can achieve high coding efficiency at high bit rates, it has the problem of being susceptible to image noise and quantization noise at low bit rates, and has not achieved sufficient performance.

[0007]

[Problems to be Solved by the Invention] As mentioned above, with conventional image coding methods, high coding efficiency can be obtained at high bit rates, but at low bit rates, prediction errors are large due to image noise and quantization noise. Therefore, there was a problem that the encoding efficiency decreased and the hardware became larger.

The present invention was made to solve the above-mentioned problems, and uses relatively simple hardware.
It is an object of the present invention to provide an image encoding device that can obtain high encoding efficiency even at a low bit rate.

[0009]

[Means for Solving the Problems] In order to solve the above problems, an image encoding device according to one aspect of the present invention includes block dividing means for dividing an input image signal into a plurality of blocks each consisting of a plurality of pixels; prediction means for obtaining a predicted value by step-by-step predicting a first pixel value to be encoded in the block of interest using a second pixel value that has already been encoded;
a prediction error calculation means for obtaining a prediction error of the predicted value obtained by the prediction means; a quantization means for obtaining a quantized value by quantizing the prediction error obtained by the prediction error calculation means;
an encoding means for encoding the quantized value obtained by the quantizing means; and an encoding means for correcting the second pixel value used in the prediction means when the quantized value obtained by the quantizing means is equal to or greater than a predetermined value. and a correction means.

An image encoding apparatus according to another aspect of the present invention further includes block dividing means for dividing an input image signal into a plurality of blocks each consisting of a plurality of pixels, and a block dividing means that divides an input image signal into a plurality of blocks each consisting of a plurality of pixels, and a block dividing means for dividing an input image signal into a plurality of blocks each consisting of a plurality of pixels. a first prediction means that obtains a first predicted value by predicting using the converted pixel values;
The first step is to obtain the prediction error of the predicted value obtained by the prediction means.
a prediction error calculation means for calculating a prediction error obtained by the first prediction means; a scalar quantization means for scalar quantizing the prediction error to obtain a quantized value when the prediction error obtained by the first prediction means is a predetermined value or more; A second method that predicts the pixel value of from the pixel value adjacent to the block of interest and the lower right pixel value within the block of interest.
a second prediction error calculation means for calculating a prediction error of the predicted value obtained by the second prediction means; It is characterized by comprising vector quantization means for vector quantizing errors to obtain quantized values, scalar quantization means, and encoding means for encoding the quantized values obtained by the vector quantization means. .

[0011]

[Operation] In the present invention, when the quantization value obtained by the quantization means is greater than or equal to a predetermined value, that is, the second
If the difference between the pixel value of the block and the first pixel value in the block is large, there is a possibility that the second pixel value used for prediction contains a large amount of quantization error. Correct the pixel value of. This reduces the prediction error and also reduces the quantization error. As a result, image noise and quantization noise are not encoded, and encoding efficiency is improved.

[0012] In this case, if the region in which the quantization width or quantization value is zero during quantization is increased at each prediction stage, the amount of code for prediction errors is reduced, so that the coding efficiency is further improved. .

Furthermore, in the present invention, the lower right pixel value in the block of interest is predicted from the already encoded pixel value adjacent to the block of interest, and when the prediction error is large, the prediction error is scalar quantized and encoded, Next, pixel values other than the lower right of the block of interest are predicted from pixel values adjacent to the block of interest and pixel values of the lower right of the block of interest, and when the prediction error is large, vector quantization and encoding are performed. When predicting the pixel value at the bottom right of a block, you are predicting spatially distant points, which often results in low prediction accuracy and large prediction errors, so it is necessary to use a large number of quantization bits. However, since both sides of pixel values other than the lower right in the block have already been encoded, the prediction accuracy is high and the prediction error is small. Therefore, the number of quantization bits can be reduced for pixel values other than the lower right in the block, and the number of pixels with sufficiently small prediction errors that do not need to be coded increases, thereby improving coding efficiency. In this case, if quantization is performed only when the prediction error is greater than or equal to the threshold, the influence of noise in the input image will be prevented and the number of pixels to be quantized will be reduced.
Encoding efficiency is further improved.

[0014]

Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing one embodiment of the present invention.

This image encoding device is composed of a control circuit 2, a memory 3, a subtracter 4, a quantizer 5, an adder 6, a predictor 7, a correction circuit 8, a memory 9, and an encoder 10. The image signal input from the input terminal 1 is written to the memory 3 for one screen, divided into multiple blocks, read out one block at a time in the coding order, and then subjected to extrapolation-interpolation predictive coding as shown below. be done.

FIG. 2 shows an encoding procedure for extrapolation-interpolation predictive encoding when the block size is 8×8. In FIG. 2, ● represents a pixel that has already been encoded, and ◯ represents a pixel that is about to be encoded.

First, in the first step shown in FIG. 2(a), the pixel value of pixel X0 in the block of interest is extracted using the pixel values of already encoded pixels A, B, and C adjacent to that block. Predict by insertion. As a method of extrapolation prediction, if the pixel value of each pixel is represented by the same symbols A, B, and C as pixels, (1) Pixel values A, B, and C are multiplied by prediction coefficients a, b, and c, respectively, and added. A method of using the calculated values a, A+b, B+c, and C as predicted values (2) A method of adaptively changing the prediction coefficient by comparing |A-B| and |C-B| (3) A method of using X0 of A and C The predicted value is the one that is closer to
There is a method of adding additional information indicating which one is selected. If the block of interest is at the top or left edge of the screen, the median of the dynamic range of pixel values is set to A, B, C.
Let the pixel value be The difference between the predicted value and the actual pixel value thus obtained, that is, the prediction error, is quantized and encoded.

Next, in the second stage shown in FIG. 2(b), interpolation coding is performed using the lower right pixel of the already encoded block and the already encoded pixels adjacent to that block. The prediction formula for this interpolation coding is

PX1=k4H(A+X0)/2PX2
=k4V(C+X0)/2 PX3=k4H(D+X2)/4+k4V(E+X1
)/4. Here, k4H is 4 in the horizontal direction.
It is a prediction coefficient when predicting from pixels far apart, and k
4V is a prediction coefficient when making predictions from pixels that are four pixels apart in the vertical direction. Note that X3 is predicted after the encoding of X1 and X2 is completed.

Interpolation predictive coding is similarly performed in the third and fourth stages shown in FIGS. 2(c) and 2(d). When the block of interest is at the top or left end of the screen, there are no already encoded pixels in the position adjacent to the block of interest, so extrapolation prediction is performed using only the pixels that have already been encoded in the block of interest.

The optimal prediction coefficient kn for interpolation prediction using the pixel values of pixels n pixels apart is kn, where the correlation coefficient is ρ.
=2ρn/(1+ρ2n). For example, when the correlation coefficient is 0.95, the prediction coefficient is 0.999, and even if the prediction coefficient is 1, a sufficiently high prediction accuracy can be obtained.

Returning to FIG. 1, the explanation will be returned to FIG. A prediction error is calculated by subtracting the pixel value read from the memory 3 and the predicted value from the predictor 7 by the subtracter 4. This prediction error is quantized by a quantizer 5. Here, the quantizer 5 is a linear quantizer that has a region (hereinafter referred to as a dead zone) where the output (quantized value) is zero for inputs (prediction errors) below the threshold TH, as shown in FIG. Consider a quantizer.

The prediction error obtained by the subtracter 4 is transmitted to the quantizer 5.
If the pixel does not fall within the dead zone, the predicted value from the predictor 7 is considered to be a significant pixel, and the quantized value obtained by the quantizer 5 is transferred to the subtracter 6 and encoder 10.

The predictor 7 performs prediction using pixels that have already been encoded, as shown in FIG. 4, for example. In FIG. 4, ● indicates the input pixel value to be encoded, ○ indicates the pixel value that has already been encoded, and ◎ indicates the predicted value of the center pixel. Since the pixel values that have already been encoded include quantization errors, there are cases where the input pixel values are shifted above and below the input pixel values. It's possible to think so. Among these, in the cases of FIGS. 4(a), 4(b), and 4(c), the prediction error is small and it is difficult to become a significant pixel. but,
When both encoded pixel values are shifted downward as shown in FIG. 4(d), the prediction error becomes large and the pixel is likely to become significant. In FIG. 4, the case where the input image signal is convex upward is considered, but if the input image signal is convex downward, significant pixels are likely to occur when both already encoded pixel values are shifted upward.

Therefore, when a significant pixel occurs, it can be said that there is a high possibility that the pixel values used for prediction are shifted in the same direction. Therefore, when a significant pixel occurs, the quantization error can be reduced by correcting the already encoded pixel value used for prediction in the same direction as the quantized value in the correction circuit 8 described later. .

The quantizer 5 is controlled by the control circuit 2 so that the quantization width or dead zone becomes larger at each stage of prediction. This allows the number of significant pixels to be deleted. However, if the quantization width or dead zone is made too large, the S/N of the decoded image will decrease, so it is necessary to find an appropriate enlargement ratio for the quantization width or dead zone.

FIG. 5 shows an example of the quantization width and expansion rate α at the final stage. In FIG. 5, the horizontal axis is the basic quantization step size, and the vertical axis is the expansion rate α of the quantization width at the final stage. This is an evaluation image of "GIRL" and "AERIA".
This figure is a plot of the optimal enlargement ratio α when ``L'', ``Carnation'', and ``Switzerland Mountain Village'' were prepared and encoded with 0.5 bits, 1 bit, and 2 bits, respectively. If this relationship is approximated by a straight line, and the basic quantization width is g and the final stage quantization width is gE, then gE = α·g=(1+g/27)g. Note that even if the quantization width or dead zone is made larger only in the final stage of prediction, it is possible to suppress the generation of significant pixels and suppress a decrease in coding efficiency.

The output of the quantizer 5 is encoded by an encoder 10 only for significant pixels whose quantization value is not 0 together with significant pixel position information. If the quantization value is not 0, the correction circuit 8 operates, and the pixel value used for prediction is input from the predictor 7 to the correction circuit 8, and after being corrected as described above using the quantization value,
It is returned to memory 9. The corrected pixel value XH is determined by the following equation. XH =X+a・g・sgn(Q)

Here, X is the pixel value used for prediction, a is the correction amount, g is the quantization width, and sgn(Q) is a sign indicating the sign of the quantization value Q. Note that the quantization width g may be expanded as described above. The correction amount a is determined by the correction circuit 8. When the quantizer 5 with a dead zone is used to quantize the prediction error, the quantization error is not corrected until a significant pixel occurs, which may cause bias in the quantized pixels. By predicting this bias in the correction circuit 8 and correcting the quantization value, it is possible to improve the S/N ratio and eliminate the number of significant pixels.

FIG. 6 shows the relationship between the correction amount a and the quantization width g. This is the pixel "GIRL""AERIAL""
0 each for ``Carnation'' and ``Switzerland Mountain Village.''
．． This is a plot of the optimal correction amount a when encoding with 5 bits, 1.0 bits, and 2.0 bits. From this, the correction amount a becomes a=0.15g. Therefore a is 0
．． About 2 is good.

[0031] In extrapolation-interpolation predictive coding, since the prediction error distribution differs at each stage of the encoding procedure, it is better to perform encoding using a different variable length encoding table at each stage. however,
Since the number of encoded pixels in the first and second stages is small, it does not have a large effect on the overall statistics, so
Even if the same encoding table is used for encoding up to the stage, the encoding efficiency does not change much. Alternatively, a plurality of sets of encoding tables may be prepared, and the encoding table with the smallest code length per block may be selected to perform encoding. At this time, the number of the selected encoding table is also encoded and transmitted. The quantized prediction error and predicted value are input to an adder 6 and locally decoded. The decoded signal is written to memory 9 and used for the next prediction.

The control circuit 2 controls the read address of the memory 3, switches the prediction coefficient of the predictor 7,
Control is performed to switch the encoding table of the encoder 10. In the encoder 10, the output signal from the quantizer 5 is subjected to, for example, Huffman encoding. This encoded signal is output from the output terminal 11 to the transmission path.

FIG. 7 shows the encoding bit rate and the S/N of decoded images when the above-mentioned quantization width is changed and the quantization error is corrected (the present invention) and when the quantization error is not corrected (the conventional example).
The results of investigating the relationship between In addition, the image is “GIRL
” “Carnation (ITE1)” “Swiss Mountain Village (I
TE4), and a linear quantizer in which one side of the dead zone is the same as the quantization width was used as the quantizer 5. The Huffman code by the encoder 10 was optimized depending on the image. As shown in FIG. 7, according to the present invention, the S/N at low bit rates is up to 1.5 d compared to the conventional technology.
It can be seen that B has also improved.

[0034] FIG. 8 also shows the difference between the present invention and the conventional AD
Regarding the CT (adaptive discrete cosine transform) method, the results of similarly examining the relationship between the encoding bit rate and the S/N of decoded images are shown. In the ADCT method, the Huffman table used in Huffman encoding by the encoder 10 is optimized depending on the image, the quantization table is made flat, and the bit rate is adjusted according to the quantization width.

In image encoding according to the present invention, the quantization width is changed stepwise and the quantization error is corrected. The images used are "GIRL", "Carnation", and "Swiss Mountain Village", and the image size of "GIRL" is 256×
256, "carnation" and "mountain village in Switzerland" are 704
×480.

As shown in FIG. 8, according to the present invention, almost the same performance as the ADCT method can be obtained by extrapolation-interpolation predictive coding with a small amount of calculation, and it is better than when using ADCT, vector quantization, etc. , the hardware can be made smaller. FIG. 9 is a block diagram showing the basic configuration of an image encoding device according to another embodiment of the present invention, and FIG. 10 is a diagram for explaining its operation.

This image encoding device includes a control circuit 22, a memory 23, a subtracter 24, a switch 25, a scalar quantizer 26, a vector quantizer 27, an adder 28, and a predictor 29.
, memory 30 and encoder 31.

In FIG. 9, the image signal inputted from the input terminal 21 is written to the memory 23 for one screen, divided into a plurality of blocks, and sequentially read out from block 1 at the upper left as shown in FIG. , quantized and encoded. Assume that the block size is 2×2, and the pixel values of the block to be currently encoded are X0 to X3. First, the pixel value X0 at the bottom right of the block is already encoded (
In this case, from block 7 C, block 8 D, E, block 12 A, B), for example, XP0=A+E-C

...(1)

Predict XP0 according to Each pixel value of A, B, C, D, and E is stored in a memory 30, and using the contents of this memory 30, the predictor 29
The calculation of equation (1) is then performed.

Next, the switch 25 is connected to the scalar quantizer 26.
A scalar quantizer 26 quantizes X0 −XP0, and an encoder 31 encodes the quantized value Q(X0 −XP0). Then, in the adder 28, the predicted value XP0 obtained by the predictor 29 is added to the quantized value obtained by the scalar quantizer 26, and the locally decoded signal XL0 shown in the following equation (2) XL0=XP0+Q( X0 - XP0)

...(2)

##EQU1## is obtained and written into the memory 30. If |X0 - XP0| is smaller than the threshold TH0, neither quantization nor encoding is performed. At this time, the local decoded signal XL0 remains the predicted value, XL0=XP0, and is written into the memory 30. The encoder 31 compresses the data using, for example, Huffman encoding.

FIG. 11 shows the scalar quantizer 2 in FIG.
FIG. 6 is a diagram showing a specific example of No. 6; When the lower right pixel data (X0 - XP0) to be quantized is input from the input terminal 40, the absolute value circuit 42 obtains the absolute value |X0 - XP0|. On the other hand, a preset threshold TH0 is input to the subtracter 43 from the input terminal 41, and H0 = |X0 −XP0 | −TH0

...(3) is output.

The determination circuit 44 determines the sign of the output H0 of the subtracter 43. If it is negative, the switch 45 is opened, and if it is positive, the switch 45 is closed. The output H0 is quantized.

The quantization characteristic of the quantizer 46 is, for example, a nonlinear quantization characteristic as shown in FIG. 12 here, but it may be a linear quantization characteristic. The adder 47 adds the threshold TH0 to the quantized value obtained by the quantizer 46, and the coding circuit 48 adds the same code as the input image signal to the input terminal 40, and then the quantized value Q(X0 -XP0) is output from the output terminal 49. Therefore, the scalar quantizer 2
The input/output characteristics of 6 are as shown in FIG.

On the decoding device side, since the value of the threshold TH0 must be known, this value is encoded first. If this threshold value is completely fixed regardless of the input image, by writing the input/output characteristics as shown in FIG.
It can be achieved with just one piece.

Although an example of a quantizer that uses the determination circuit 44 to determine whether or not to perform quantization has been described here, the quantizer has a configuration that always performs quantization of X0 - XP0. may also be used.

Next, pixel values X1 and X2 other than X0
, X3 is predicted. Predicted value XP1 of these pixels,
XP2 and XP3 are obtained by calculating, for example, the following equations (4) to (6). XP1=(A+XL0)/2

...(4) XP2=(E+XL0)/2

...(5) XP3=(C+XL0)
/2
...(6)

[0048] And,
These predicted values XP1, XP2, XP3 and X1, X2
, X3 and the difference value X1 - XP1, X2 - XP2,
Residual signal vector e=(e
1, e2, e3) are created. ei =Xi −XP1

(7) Next, the residual signal vector e is quantized in the vector quantizer 27.

FIG. 14 is a diagram showing a specific example of the vector quantizer 27. When the residual signal vector e to be quantized is input from the input terminal 50, M code vectors c(m) (m=1
~M) are sequentially input to the inner product calculation circuit 51. The code vector c(m) (m=1 to M) is composed of elements with the same number of dimensions as the residual signal vector, and in this example,
c(m) = (C1 (m), C2 (m), C
3 (m) ) …(8)
It is expressed as Here, the amplitude of each code vector is
c(m) ・ct(m)=1
…
It is normalized in advance so that it becomes (9). t represents the transpose of the matrix.

[0050] The code number is the code vector c(m)
The vector γ(m) ・c(
m) and the residual signal vector e are selected from m=1 to M so that the distance D shown in the following equation is minimized. D=‖e−γ(m)・c(m) ‖
…(10
) For example, if we use Euclidean distance as the distance, D is

D=(e-γ(m)・c(m))(e-γ(
m)・c(m) )t

...(11). The minimum value of the distance D when the optimum gain γ(m) is given by this equation (11) is given by the following equation. Dmin (m) = e・et −{γ(m)
}2...(1
2) Here, the gain γ(m) is calculated by the inner product calculation circuit 51 using the following formula: γ(m) = e・ct(m)

...(13) is calculated.

The code number k that minimizes the left side of equation (12) is found by selecting the code number m=1 to M that maximizes the absolute value of gain γ(m) in equation (13). It will be done.

The absolute value circuit 53 determines the absolute value of the gain γ(m). The code circuit 55 determines the sign of the gain γ(m) and outputs the code of the gain γ(k) corresponding to the code number k searched by the maximum value search circuit 54 to the quantizer 56. The quantizer 56 calculates the sign SGN(
Enter γ(k)) and absolute value ABS(γ(k)),
The quantized value Q(γ(k)) of the optimal gain is output.

The multiplication circuit 57 outputs a vector obtained by multiplying the code number k by the code vector c(k), that is, a quantized value eQ of the residual signal vector e. eQ = Q(γ(k))・c(k)
…(
14) Therefore, the local decoded signal XLi of the current block is
Xi ~=XPi+eLi, i=1,2,
3...(15).

Here, the vector quantizer 27 shown in FIG.
4, but similar to the scalar quantizer 26,
If a vector quantizer configured to not perform vector quantization when the residual signal vector to be quantized that is input to the quantizer is close to a zero vector, encoding efficiency can be further improved.

FIG. 15 shows an example of this vector quantizer. Residual signal vector e input from input terminal 60
is input to the determination circuit 63. The determination circuit 63 calculates the power e·et of the residual signal vector, connects the switch 61 only when this is larger than a predetermined threshold Te, inputs the residual signal vector e to the vector quantizer 62, and converts the residual signal vector e into the vector quantizer 62. Perform quantization. Here, the vector quantizer 62 in FIG. 15 has the same configuration as the vector quantizer in FIG. 14.

In this method, information as to which pixel's pixel value was encoded must also be encoded. As for this encoding method, for example, a block with one or more encoded pixels is treated as a significant block, a block with no encoded pixels is treated as an invalid block, a significant block is set as "1", an unsigned block is set as "0", and the code is set as "1". The sequence can be run-length encoded. This significant block determination is based only on whether or not X0 is encoded; if X0 is not encoded, then
1 to X3 may also not be encoded. In the latter method, TH0 = 0 and X0 of all blocks
When encoding the block, there is no need to encode the significance determination information of the block. In the case of a significant block, information as to which pixel was encoded must also be encoded, but this information can be compressed by allocating one bit to each pixel and using Huffman encoding or the like.

In the above example, the block size was 2×2, but an example of encoding when the block size is 4×4 pixels will be explained below with reference to FIG. First, as shown in FIG. 16A, the prediction residual (represented by ◯) X0 - XP0 of the 4×4 lower right pixel X0 is encoded to obtain XL0.

Next, as shown in FIG. 16(b), 2×2
Let the prediction residuals (○) of the lower right three pixels other than those sent in the previous stage (represented by become

Next, as shown in Figure (c), three-dimensional prediction residual signal vectors e2, e3, e4, e5 are obtained from the prediction residuals (○) of pixels other than those sent in the previous stage (●).
are created, and vector quantization is performed if the powers of these vectors are greater than or equal to the respective thresholds Te2, Te3, Te4, and Te5.

Another possible encoding method is the method shown in FIG. 17, for example. In this method, first, Figure 1
7(a), the prediction residual X0-XP0 of the 4×4 lower right pixel X0 is encoded to obtain XL0.

Next, as shown in FIG. 17(b), pixel A
, C, D, E, XL0, the predicted residual signal vector e
1 is obtained, and if this power is greater than or equal to the threshold Te1, vector quantization is performed.

Next, as shown in FIG. 17(c), the prediction of the three pixels in the upper left of the pixels other than those sent in the previous stage (●) is performed using the five adjacent ● pixels, and A residual signal vector e2 is obtained, and if this power is greater than or equal to the threshold Te2, vector quantization is performed. Next, as shown in FIG. 17(d), among the pixels other than those sent in the previous stage (●),
For the pixel indicated by , prediction is performed using five adjacent ● pixels to obtain residual signal vectors e3 and e4, and if the powers are greater than or equal to the respective thresholds Te3 and Te4, vector quantization is performed.

Next, as shown in FIG. 17(e), for the three pixels indicated by ○ other than those sent in the previous stage (●),
Prediction is performed using five adjacent ● pixels to obtain a residual signal vector e5, and if this power is greater than or equal to the threshold Te5, vector quantization is performed.

[0065] When vector quantization is performed using such a procedure, the pixel value at the position of ○ can be predicted using the pixel value at the position of ●, which is closer than that in the case of Fig. 16, so the prediction performance is improved. , an improvement in encoding efficiency can be expected. Figure 16(a)
) and the predicted value XP0 of X0 in FIG. 17(a) can be, for example.

[0066] Here, we have explained only the case of intra-frame prediction as a prediction method, but when applied to moving images, coding is performed separately for inter-frame prediction and intra-frame prediction, and if the prediction method is greater than or equal to a threshold, The code with the smaller number of significant vectors may be transmitted. In the case of inter-frame prediction, the motion-compensated corresponding pixel of one frame before is used as the predicted value.

In the above example, the case where the block size N×M is an even number was explained, but for example, as shown in FIG.
It is also applicable to the case of ×3. Although X0 is the same as above, for example, the methods of determining X1 and X4 are different. In this case, the ratio of X1 and X4 may be changed as shown below. The same applies to X2, X3, etc.

[0068]

Effects of the Invention As explained above, according to the present invention, image encoding is improved by correcting already encoded pixel values or by increasing the quantization width or dead zone at each stage of prediction. This improves the disadvantage of being susceptible to image noise and quantization noise in image encoding devices that use extrapolation-interpolation predictive encoding, which is one of the methods, and improves encoding efficiency at low bit rates. .

Furthermore, according to the present invention, within a small block of, for example, 2 pixels x 2 pixels, the lower right pixel in the block is first predicted, scalar quantized and encoded, and then other points are interpolated. By predicting, vector quantizing, and encoding, the prediction error is reduced for pixels other than the bottom right pixel in the block, so vector quantization is more efficient than vector quantization using conventional extrapolation prediction. be able to. In this case, by performing scalar quantization and vector quantization only when the prediction error is equal to or greater than a predetermined threshold, it is possible to prevent an increase in entropy due to noise in the input image, and improve coding efficiency.

[Brief explanation of the drawing]

[Fig. 1] Block diagram of an image encoding device according to an embodiment of the present invention

[Figure 2] Diagram showing the procedure of extrapolation-interpolation predictive coding in the same embodiment

[Figure 3] Diagram showing the input/output characteristics of the quantizer in Figure 1

[Figure 4] Diagram explaining the method of correcting quantized values in the same embodiment

[Figure 5] Diagram showing the relationship between the final stage quantization value and expansion rate in the same example

[Figure 6] Diagram showing the relationship between the quantization value and its correction amount in the same example

[Figure 7] A diagram showing the relationship between encoding bit rate and S/N in the same embodiment and the conventional extrapolation-interpolation predictive encoding method.

[Fig. 8] A diagram showing the relationship between encoding bit rate and S/N in the same embodiment and the conventional ADCT method.

[Figure 9
] Block diagram of an image encoding device according to another embodiment of the present invention

[Figure 10] Diagram for explaining the operation of the same embodiment

[Figure 11] Block diagram showing the configuration of the scalar quantizer in Figure 9

[Figure 12] Diagram showing the input/output characteristics of the quantizer in Figure 11

[Figure 13] Diagram showing the input/output characteristics of the scalar quantizer in Figure 9

[Figure 15] Block diagram showing the configuration of the vector quantizer in Figure 9

[FIG. 16] Block diagram showing another configuration of the vector quantizer in FIG. 9

[Figure 17] Diagram for explaining the method of calculating predicted values in the same example

[Figure 18] Diagram for explaining the method of calculating predicted values in the same example

[0070]

[Explanation of symbols]

1...Input terminal
2...Control circuit 3, 9...Memory
4...Subtractor 5
…quantizer
6...Adder 7...Predictor
8...Correction circuit 10...Encoder
11...Output terminal 21...Input terminal
22...Control circuit 23...Memory
24...Subtractor 25...Switch
26... Scalar quantizer 27... Vector quantizer
28...Adder 29...Predictor
30...Memory 31...Encoder
32...Output terminal

Claims (6)

[Claims] 1. Block dividing means for dividing an input image signal into a plurality of blocks each consisting of a plurality of pixels; and a second pixel that has already been encoded a first pixel value to be encoded in the block of interest. A prediction means for obtaining a predicted value by step-by-step prediction using the predicted value, a prediction error calculation means for obtaining a prediction error of the predicted value obtained by the prediction means, and a prediction error obtained by the prediction error calculation means. quantization means for quantizing to obtain a quantized value; encoding means for encoding the quantized value obtained by the quantization means; and when the quantized value obtained by the quantization means is greater than or equal to a predetermined value. and a correction means for correcting the second pixel value used in the prediction means. 2. The image encoding device according to claim 1, wherein the prediction means uses a second pixel value in a block adjacent to the block of interest or within the block of interest as the second pixel value. 3. The image encoding apparatus according to claim 1, wherein the correction means corrects the second pixel value when the quantization value is other than zero. 4. The image encoding apparatus according to claim 1, wherein said quantizing means increases a region in which a quantization width or a quantization value becomes zero at each stage of said predicting means. 5. Image encoding according to claim 1, wherein the correction means corrects the second pixel value used by the prediction means in the same direction as the quantized value of the prediction error. Device. 6. Block dividing means for dividing an input image signal into a plurality of blocks each consisting of a plurality of pixels; a first prediction means for obtaining a predicted value of 1; a first prediction error calculation means for obtaining a prediction error of the predicted value obtained by the first prediction means; and a prediction obtained by the first prediction means. scalar quantization means for scalar quantizing the prediction error to obtain a quantized value when the error is greater than or equal to a predetermined value; a second prediction means for predicting from the lower right pixel value of the second prediction means, a second prediction error calculation means for obtaining a prediction error of the predicted value obtained by the second prediction means, and the second prediction means. vector quantization means for vector quantizing the prediction error to obtain a quantized value when the obtained prediction error is greater than or equal to a predetermined value; An image encoding device comprising: encoding means for encoding an image.