JPH11127439A - Image encoding device, its method, image decoding device and its method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce work memory or to make it unnecessary by outputting new high pass components and producing a predictive value by appropriately using extrapolation estimation only from the past and interpolation estimation from the past and future. SOLUTION: A predictive value generating part 73d of a vertical inverse converting part 73 produces a predictive value having a limiter from low pass coefficients 77 and 80. An adding part 73c adds the predictive value having a limiter and an inverse quantization coefficient, and an operating part 73d operates the next low pass component from acquired high pass components and outputs operating part output signals VACMD0 to VACM0D7. Also, a register part 73e collectively stores them when all data that are acquired are gathered and outputs each output signal IVRD0 to IVRD7. Further, a selector part 73f discharges stored data as an inverse conversion output RMIN one by one. In this way, new high pass components are acquired and the new high pass components are outputted and also, predictive value is generated by appropriately using extrapolation estimation only from the past and interpolation estimation from the past and future.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image encoding apparatus and method using wavelet transform and inverse transform used in highly efficient encoding of image signals and the like, and an improvement of an image decoding apparatus and method.

[0002]

2. Description of the Related Art Conventionally, for example, in a so-called image signal transmission system for transmitting an image signal to a remote place, such as a video conference system, and in an apparatus for digitizing an image signal and recording and reproducing it on a compact disk or the like, In order to efficiently use transmission paths and recording media, significant information is efficiently coded by utilizing the correlation of digitized image signals, thereby reducing the amount of transmitted information and recorded information, and improving transmission efficiency and recording. It is designed to increase efficiency.

[0003] Encoding methods using correlation of image signals include, for example, predictive encoding methods, orthogonal transform encoding methods such as DCT (discrete cosine transform), subband encoding methods, and wavelet transform methods. For example, a method of dividing an image signal into a plurality of components, quantizing the image signal, and transmitting the image signal is used.

[0004] Among them, the predictive coding method is easy to implement and suitable for coding with a relatively low compression rate. However, when the compression rate is increased, deterioration of image quality is easily detected. is there. In addition, orthogonal transform coding methods such as DCT are often used because high image quality can be obtained relatively easily with a high compression ratio. There is a drawback that obstacles caused by different degrees are easily noticeable.

In the orthogonal transform coding method using DCT, as shown in FIG. 19, a DCT 101 in the horizontal direction is performed and stored in a buffer memory 102, and then the data is stored in a vertical direction.
Perform CT103. Then, quantization 104 is performed, and variable-length coding 105 is performed and output. The large amount of calculation of the DCT causes an increase in the circuit scale when implemented by hardware, an increase in the calculation time when implemented by software, and the like. In particular, DCT is
A multiplier is required because it includes real number operation, and when trying to process a large-sized image at high speed, a plurality of multipliers must be provided to perform the operation in parallel, causing a problem that the circuit scale becomes large. I have.

Further, a band division encoding method such as a broad band subband encoding method including a wavelet transform method divides an image signal into a plurality of bands and then performs quantization. Obtained, and characteristic distortion is hardly noticeable. Although there is a difference in the device implementation depending on the method, a filter corresponding to the following equations (1) to (4), that is, a filter corresponding to the following equations (1) to (4), that is, a wavelet transform using a Haar basis is extremely simple. There are also methods that can be configured.

[0007]

(Equation 1)

The encoding by the wavelet transform method is as follows.
Band division is performed recursively on the low-order side band or the low-order side band, and encoding is performed in accordance with the characteristic of each band obtained as a result, that is, each coefficient representing the band. Then, in the case of compression of a digital signal, a discrete and orthogonal basis expression is obtained. When one-dimensional orthogonal wavelet transform is actually performed on a digital signal, a two-divided filter bank is used, and octave division described later is employed.

[0009] The configuration of the band splitting filter bank for the first-order orthogonal wavelet transform when using the Haar base satisfying such conditions is very simple as shown in FIG. In FIG. 20, a signal input from an input terminal 111 is supplied to a low-pass filter 112 and a high-pass filter 113. In the low-pass filter 112, a low-order coefficient is extracted by performing a filtering process using the transfer function H ₀ shown in Expression (1), and the obtained low-order coefficient is down-sampled by the down-sampler 114 to 2: 1 ( = Thinned out) and output to the output terminal 116.
Similarly, in the high-pass filter 113, a high-order coefficient is extracted by performing a filtering process using the transfer function H ₁ shown in the equation (2), and the high-order coefficient is thinned out by the down sampler 115 by 2: 1 and output to the output terminal 117. You.

[0010] On the other hand, the band synthesis filter bank for the inverse transform used when decoding the coded one has a simple configuration as shown in FIG. The signals are input from terminals 121 and 122, respectively, and 0 coefficients are alternately interpolated by upsamplers 123 and 124 to double the sample. On the low-order coefficient side, the transfer function H shown in Expression (3) is obtained by the low-pass filter 125.
_The filter operation by ₃ is performed to interpolate the 0 coefficient, and on the high-order coefficient side, the 0 coefficient is interpolated by the high-pass filter 126 by the filter operation by the transfer function H ₄ shown in Expression (4). Then, the two are added by the adder 127, reconstructed into the original signal, and output to the output terminal 128.

In the case of an image signal, the above-described conversion is performed in the horizontal and vertical directions in consideration of the two-dimensional structure of the image, and the conversion is performed by octave division. That is, as shown in FIG. 22, after dividing into two in the vertical direction, dividing each into two in the horizontal direction, dividing the entire band into four, and then recursively dividing only the lower band where power is most concentrated. Then, conversion that finally divides into a plurality of bands is common.

As a technique in which the above-described wavelet transform is specifically applied to image coding, for example, there is a method disclosed in Japanese Patent Application Laid-Open No. 6-292185. This technique performs a two-dimensional wavelet transform on an image signal by the above-described method, and quantizes the obtained coefficients into coarse coefficients on the high-order side and finely quantizes the coefficients on the low-order side in consideration of human visual characteristics. The variable coefficient is subjected to variable length coding using the run length code or the Huffman code. In addition, Japanese Unexamined Patent Application Publication No.
In the technique of -292185, an attempt is made to improve block-like distortion or mosaic distortion of a reconstructed image, which is a problem of wavelet transform using a Haar basis, and a higher-order coefficient corresponding to a change point of a lower-band basis function is intended. The method of quantizing more finely than other coefficients is introduced.

As described above, the conventional image coding apparatus and image decoding apparatus using the wavelet transform generally quantize the low-order coefficients finely and consider the high-order coefficients in consideration of human visual characteristics. The coefficient on the side is coarsely quantized. As a result, the information amount of data to be transmitted is reduced as a whole. However, when quantization is performed in this way, when the compression rate is increased, or when an image that is difficult to compress due to including many high-frequency components is input, the frequency at which the coefficients are coarsely quantized increases. As a result, a block-like or mosaic-like distortion is easily observed, and there is a problem that the image quality is visually deteriorated.

[0014] This mosaic distortion is liable to be concentrated on the change point of the low-pass side basis function in the band division using, for example, the Haar basis, and it is difficult to improve the mosaic distortion with a normal low-pass filter. This is an obstacle to the application of the Haar transform, which has an excellent feature of easy conversion, to an image encoding device and an image decoding device.

The technique disclosed in Japanese Patent Application Laid-Open No. 6-292185 has been made to improve this point.
Finely quantizing the change point of the low-frequency basis function, that is, the higher-order coefficient corresponding to the block boundary, increases the amount of generated code, and even if block-like distortion or mosaic-like distortion is improved, Encoding efficiency, that is, the compression ratio may be reduced.

[0016]

In order to solve such a problem, the inventor of the present invention has previously described a simple configuration similar to the conversion using the Haar base, and has no block-like or mosaic-like distortion. And an application for a wavelet transform apparatus and method and a wavelet inverse transform apparatus and method for improving coding efficiency and decoding efficiency at the same time (Japanese Patent Application No. 8-50634).

The wavelet transform apparatus and method, and the wavelet inverse transform apparatus and method,
A predicted value for predicting a high-frequency component is generated using the low-frequency component. Then, the predicted value is subtracted from the high frequency component to obtain a new high frequency component, or the predicted value is added to the high frequency component to obtain a new high frequency component. Therefore, the power of the high frequency component is reduced, the entropy of the high frequency component can be significantly reduced, and the encoding efficiency and the decoding efficiency are improved.

On the other hand, in the case of using the conventional wavelet transform and inverse transform, the power of the high frequency component, that is, the power of the high order coefficient tends to be large. For example, a higher-order coefficient in the case of using the Haar basis is obtained by taking a difference between adjacent pixel pairs as indicated by the transfer function of Expression (2). Therefore, the loss of higher-order coefficients due to quantization or the like means that the pixel pairs are reconstructed to the same level, especially in the low-frequency region where the shading of the image changes gradually. Causes distortion.

However, in the low frequency region, the difference between the high frequency component, ie, the high order coefficient, and the low frequency component, ie, the low order coefficient, has a high correlation. For example, as shown in FIG. 23, the difference value (higher order coefficient) between pixel C and pixel D is the average value (lower order coefficient) of pixel A and pixel B and the average value (lower order coefficient) of pixel E and pixel F. ) Can easily be inferred to be approximately equal to the difference value of ４.

The invention filed earlier by the present inventor has focused on this point, and performs filter processing such as taking the difference between low-order coefficients on the division side, and raises the value multiplied by an appropriate prediction coefficient to a high value. The predicted value of the next coefficient is used, and the predicted value is subtracted from the higher coefficient to perform encoding processing. On the other hand, in decoding, by adding the same predicted value to the higher-order coefficients during synthesis, even if the higher-order coefficients are lost due to quantization or the like, high-frequency components such as images can be reproduced with lower-order coefficients. I have. Therefore, in the case of an image signal, a good image free from block-like or mosaic-like distortion is obtained.

When higher-order coefficients are not lost due to quantization or the like, those higher-order coefficients are encoded as significant coefficients. However, in order to improve the coding efficiency, it is important to set as many high-order coefficients as possible to 0 and to make the entropy of the whole coefficients smaller. On the other hand, generally, an image has a larger area in a low-frequency region that changes smoothly than in a high-frequency region such as an outline. According to the invention filed earlier, by subtracting the predicted value generated from the low-order coefficient from the high-order coefficient, the high-order coefficient can be converted to 0 or a level that can be regarded as 0 in many cases. as a result,
The entropy of the higher-order coefficient can be made smaller, and the coding efficiency and the decoding efficiency can be greatly improved as a whole.

Although the invention filed earlier has been considerably improved in terms of image quality, it is necessary to sequentially convert high-order coefficients to low-order coefficients over the entire screen. Work memory is required. In other words, a memory for storing the results during the operation is required for the entire screen. In addition, since access to a large-capacity work memory frequently occurs in the course of operation, when processing is performed by software, there is a problem that the hit rate of the cache memory decreases and the processing speed decreases.

In the inventions filed earlier, the outlines of the wavelet transform device and the inverse wavelet transform device have been described. However, a specific hardware configuration for efficiently processing an image is not described in detail.

An object of the present invention is to provide an image encoding apparatus and method, and an image decoding apparatus and method, in which a work memory is significantly reduced or completely eliminated. It is another object of the present invention to provide an image decoding apparatus and method that embody the prediction method and operate efficiently.

[0025]

In order to achieve the above object, the invention according to claim 1 comprises a low-frequency component extracting means for extracting a low-frequency component from an input image signal, and a high-frequency component from an image signal. A high-frequency component extraction means for extracting a high-frequency component, a prediction value generation means for generating a prediction value for predicting a high-frequency component, and a subtraction means for subtracting the prediction value from the high-frequency component, The new high-frequency component is obtained by the subtraction means, the new high-frequency component is output, and the predicted value generation means generates a predicted value by appropriately using extrapolation prediction only from the past and interpolation prediction from the past and future. ing.

As described above, the power of the high frequency component is reduced by using the predicted value, the entropy of the high frequency component can be greatly reduced, and the coding efficiency is improved. In addition, since the prediction value is predicted by extrapolation, when applied to processing of an image signal, it can be applied to encoding of a divided image instead of the entire image.

According to a second aspect of the present invention, in the image encoding apparatus according to the first aspect, an image signal is divided into predetermined blocks, and extrapolation is performed when prediction for generating a prediction value extends between blocks. Prediction is performed, and when it does not span between blocks, interpolation prediction is performed, and an upper limit and a lower limit are set for a predicted value by extrapolation prediction.

As described above, since the image signal is divided into predetermined blocks and processed, the execution speed can be improved and the circuit scale can be reduced. In addition, since the upper limit and the lower limit are set for the prediction value, it is possible to prevent a prediction error from increasing even if extrapolation prediction (one-way prediction) is adopted in prediction over blocks. As a result, the degree of power concentration on the low-frequency coefficient of the conversion result can be increased, so that the image quality is equivalent to DCT.

Further, the invention according to claim 3 is an image comprising a low-frequency component extracting step of extracting a low-frequency component from an input image signal, and a high-frequency component extracting step of extracting a high-frequency component from the image signal. In the encoding method, a prediction value generation step of generating a prediction value for predicting a high frequency component, and a subtraction step of subtracting the prediction value from the high frequency component are provided, and a new high frequency component is obtained by the subtraction step. In the prediction value generation step, a prediction value is generated by appropriately using extrapolation prediction only from the past and interpolation prediction from the past and the future.

As described above, the power of the high frequency component is reduced by using the predicted value, the entropy of the high frequency component can be greatly reduced, and the coding efficiency is improved. In addition, since the prediction value is predicted by extrapolation prediction, the prediction value can be applied to coding of a divided image instead of the entire image.

[0031] In addition, the invention according to claim 4 provides the invention according to claim 3.
In the described image encoding method, the image signal is divided into predetermined blocks, extrapolation prediction is performed when prediction for generating a prediction value extends between blocks, and interpolation prediction is performed when the prediction does not extend between blocks. In addition, an upper limit and a lower limit are set for the predicted value obtained by extrapolation prediction.

As described above, since the image signal is divided into predetermined blocks and processed, the execution speed can be improved and the circuit scale can be reduced. In addition, since the upper limit and the lower limit are set for the prediction value, it is possible to prevent a prediction error from increasing even if extrapolation prediction (one-way prediction) is adopted in prediction over blocks. As a result, the degree of power concentration on the low-frequency coefficient of the conversion result can be increased, so that the image quality is equivalent to DCT.

According to a fifth aspect of the present invention, there is provided a low-frequency component restoring means for inputting and restoring a low-frequency component of an image signal extracted by wavelet transform, and a high-frequency component of the image signal extracted by wavelet transform. An image decoding apparatus comprising: a high-frequency component restoring means for inputting and restoring; a predicted value generating means for generating a predicted value for predicting a high-frequency component from a low-frequency component; An adding means for obtaining a band component, and a synthesizing means for synthesizing the low-frequency component and the new high-frequency component, wherein the prediction value generating means appropriately performs extrapolation prediction only from the past and interpolation prediction from the past and the future. To generate forecasts.

As described above, the power of the high frequency component is reduced by using the predicted value, the entropy of the high frequency component can be greatly reduced, and the decoding efficiency is improved. Further, since the prediction value is predicted by extrapolation prediction, when the prediction value is applied to processing of an image signal, the prediction value can be applied to decoding of a divided image instead of the entire image.

Further, the invention described in claim 6 is the same as claim 5
In the described image decoding apparatus, it is assumed that an input signal is divided into predetermined blocks, extrapolation prediction is performed when prediction for generating a prediction value is between blocks, Interpolation prediction is performed, and an upper limit and a lower limit are set for the predicted value obtained by extrapolation prediction.

As described above, since the input signal is divided into predetermined blocks and processed, the execution speed can be improved and the circuit size can be reduced. In addition, since the upper limit and the lower limit are set for the prediction value, it is possible to prevent a prediction error from increasing even if extrapolation prediction (one-way prediction) is adopted in prediction over blocks. As a result, the degree of power concentration on the low-frequency coefficient of the conversion result can be increased, so that the image quality is equivalent to DCT.

According to a seventh aspect of the present invention, there is provided a low-frequency component restoring step of inputting and restoring a low-frequency component of an image signal extracted by wavelet transform, and a high-frequency component of the image signal extracted by wavelet transform. An image decoding method comprising a high-frequency component restoring step of inputting and restoring, wherein a predicted value generating step of generating a predicted value for predicting a high-frequency component from a low-frequency component, and adding the predicted value to the high-frequency component to obtain a new It has an addition step of obtaining a band component and a synthesis step of synthesizing a low band component and a new high band component, and the prediction value generation step appropriately uses extrapolation prediction from only the past and interpolation prediction from the past and the future. To generate a predicted value.

As described above, since the input signal is divided into predetermined blocks and processed, the execution speed can be improved and the circuit scale can be reduced. In addition, the power of the high-frequency component is reduced by using the predicted value, and the entropy of the high-frequency component can be significantly reduced, thereby improving the decoding efficiency. Also,
Since the prediction value is predicted by extrapolation, it can be applied to decoding of a divided image instead of the entire image.

Further, the invention according to claim 8 provides the invention according to claim 7
In the described image decoding method, the input signal is assumed to be divided into predetermined blocks, extrapolation prediction is performed when prediction for generating a prediction value extends between blocks, and Interpolation prediction is performed, and an upper limit and a lower limit are set for the predicted value obtained by extrapolation prediction.

As described above, since the input signal is divided into predetermined blocks and processed, the execution speed can be improved and the circuit size can be reduced. In addition, since the upper limit and the lower limit are set for the prediction value, it is possible to prevent a prediction error from increasing even if extrapolation prediction (one-way prediction) is adopted in prediction over blocks. As a result, the degree of power concentration on the low-frequency coefficient of the conversion result can be increased, so that the image quality is equivalent to DCT.

According to a ninth aspect of the present invention, an image signal is divided into blocks each having a predetermined number of pixels, and one of the divided image signal in one of a vertical direction and a horizontal direction is subjected to wavelet transform. 1, a buffer memory for temporarily storing the coefficients of the low-frequency component and the high-frequency component extracted by the first conversion unit, and a vertical and horizontal direction of the divided image signal read from the buffer memory. Second wavelet transform of one of the signals in the other direction
, A quantizing means for quantizing the coefficient obtained by the second converting section, and a variable-length code for assigning a code to the quantized coefficient in accordance with the magnitude of its occurrence probability and performing variable-length coding. In the image encoding apparatus including the encoding unit, at least one of the first and second conversion units includes a low-frequency component extraction unit configured to extract a low-frequency component from an input image signal; A high-frequency component extracting means for extracting the high-frequency component; a predicted value generating means for generating a predicted value for predicting the high-frequency component; and a subtraction means for subtracting the predicted value from the high-frequency component. The band component is obtained, the new high band component is output, and the prediction value generation means generates a prediction value by appropriately using extrapolation prediction only from the past and interpolation prediction from the past and the future.

As described above, in this image coding apparatus, the power of the high frequency component is reduced by employing the predicted value, the entropy of the high frequency component can be greatly reduced, and the coding efficiency is improved. Moreover, since the algorithm does not use multiplication, the circuit scale is smaller than that of DCT. Further, since the image can be processed in small block units, a work memory is unnecessary or can be greatly reduced. Further, since the prediction is corrected, the degree of power concentration on the low-frequency coefficient of the conversion result is increased, and the image quality can be achieved as high as DCT.

According to a tenth aspect of the present invention, there is provided a variable-length decoding means for dividing an input signal into blocks each having a predetermined number of pixels and performing variable-length coding, and decoding the image signals with variable length. An inverse quantization means for inversely quantizing the variable-length-decoded signal; and a first inversely inverse wavelet transform for the inversely-quantized signal and one of the vertical and horizontal signals of the image signal. A transformation unit, a buffer memory for temporarily storing the inversely transformed signal, and a second inverse transformation unit for performing a wavelet inverse transformation on a signal in the other direction of the divided image signal read from the buffer memory. In the image decoding device, at least one of the first and second inverse transform units is configured to generate a predicted value for predicting a high frequency component of the image signal from a low frequency component of the image signal. And an adder that adds the predicted value to the inversely quantized value to obtain a new high-frequency component, and an operation that decodes a value in a target direction of the image signal using the low-frequency component and the new high-frequency component Part.

As described above, in this image decoding apparatus, the power of the high frequency component is reduced by employing the predicted value, the entropy of the high frequency component can be greatly reduced, and the decoding efficiency is improved. Moreover, since the algorithm does not use multiplication, the circuit scale is smaller than that of DCT. Further, since the image can be processed in small block units, a work memory is unnecessary or can be greatly reduced.

Further, the invention according to claim 11 is the image decoding device according to claim 10, wherein the predicted value generation unit comprises:
If the predictions for generating predicted values span blocks,
Extrapolation prediction is performed only from the past, and if it does not span between blocks, interpolation prediction is performed from both the future and the past, and upper and lower limits are set for the predicted value by the extrapolation prediction.
For this reason, even if the image is processed for each block, the prediction error does not increase, and the degree of power concentration on the low-frequency coefficients of the conversion result increases, so that the image quality can be made comparable to DCT.

According to a twelfth aspect of the present invention, in the image decoding apparatus according to the tenth or eleventh aspect, at least one of the inverse transform units stores all signals in the same predetermined block obtained by the arithmetic unit. It comprises a register section and a selector for sequentially outputting each signal in the register section. As described above, since the register section and the selector exist, it is possible to sequentially send out the inverted signal.

The thirteenth aspect of the present invention provides the first aspect.
In the image decoding device described in 0, 11, or 12, the prediction value generation unit has a register that stores a low-frequency component output from the operation unit. Therefore, when it is desired to use a past output instead of a new operation unit output for generating a predicted value, it is possible to extract and use a past operation unit output from a register.

[0048]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 shows a configuration example of a band division filter bank in a wavelet transform according to the present invention. This band filter bank reduces the entire image to 1
One is divided into a plurality of blocks of 8 × 8 pixels and processed. As shown in FIG. 1, a signal input terminal 11 and a low-pass filter 12 for extracting low-order coefficients which are low-frequency components are provided. , A high-pass filter 13 for extracting a higher-order coefficient to be a high-frequency component, and a downsampler 1 for thinning out
4 and 15, a predicted value generating filter 16 for generating a predicted value of a higher-order coefficient, a register 17 for delaying the output by the delay caused by the high-pass filter 13, a subtractor 18 serving as a subtracting means, It comprises a higher order coefficient output terminal 19 and a higher order coefficient output terminal 20. The low-pass filter 12 constitutes low-frequency component extraction means, the high-pass filter 13 constitutes high-frequency component extraction means, and the predicted value generation filter 16 constitutes predicted value generation means.

Here, the low-pass filter 12 corresponds to FIG.
The digital filter is the same digital filter as the low-pass filter 112 shown in FIG. The high-pass filter 13 is also the same digital filter as the high-pass filter 113 shown in FIG. 13, and is a Haar-based filter shown in Expression (2). The downsampler 14 selects and extracts the low-order coefficients obtained by the low-pass filter 12 as appropriate.
The predicted value generation filter 16 has transfer functions H _{21 to} H ₂₄ shown in the following equations (5) to (8), and
The difference between the low-order coefficients selected in Step 4 is obtained.
Note that k in Equations (5) to (8) is a prediction coefficient, and k is 1
At this time, the linear approximation is obtained.

[0051]

(Equation 2)

The transfer functions H _{21 to} H ₂₄ have a limiter for reducing a prediction error, as described later. This limiter is a function that limits the output to a certain range.

Next, the operation will be described. The signals input from the signal input terminals 11 are applied to the low-pass filter 1 respectively.
2 is supplied to the high-pass filter 13 and divided into low-order coefficients representing low-frequency components and high-order coefficients representing high-frequency components. The low-pass filter 12 and the high-pass filter 13 perform filter processing using transfer functions as shown in Expressions (1) and (2). After the division, the low-frequency coefficients are appropriately thinned out via the downsampler 14. The low-order coefficient is output to the low-order coefficient output terminal 19 and supplied to the predicted value generation filter 16.

Then, the predicted value generation filter 16 performs a filter process using transfer functions H _{21 to} H ₂₄ shown in equations (5) to (8). Then, the output of the prediction value generation filter 16 is subtracted from the higher-order coefficient delayed by the register 17 and output to the higher-order coefficient output terminal 20.

Next, the operation of such a band division filter bank will be described with reference to a specific example shown in FIG. As shown in FIG. 3, A, B, C, D, E,
Assuming that there is a straight line connecting F, G, and H pixels,
In the low-pass filter 12, (A + B) / 2, (C +
D) / 2, (E + F) / 2 and (G + H) / 2 signals are produced. Then, this signal is appropriately thinned out by the downsampler 14, and (A + B) / 2, (C + D) / 2, (E
+ F) / 2 and (G + H) / 2 are output to the output terminal 19 and supplied to the predicted value generation filter 16.
On the other hand, in the high-pass filter 13, (AB) / 2,
(BC) / 2, (CD) / 2, (DE) / 2,
(EF) / 2, (FG) / 2 and (GH) / 2
Are generated. And this signal down sampler 1
5, thinned out as appropriate, (AB) / 2, (CD) / 2,
(EF) / 2 and (GH) / 2 are supplied to the register 17 and temporarily stored.

The signals supplied to the predicted value generating filter 16 are transferred to the predicted value generating filter 16 by transfer functions H _{21 to} H _24.
[(A + B) / 2- (C + D) / 2] × 1 /
4, [(A + B) / 2- (E + F) / 2] × １ ／,
[(C + D) / 2- (G + H) / 2] × 1/8, [(E
+ F) / 2- (G + H) / 2] .times.1 / 4 (provided that the prediction coefficient k is 1 and is not limited by the limiter). Then, these signals are converted to the previous signal (A-
B) / 2, (CD) / 2, (EF) / 2, (G-
H) / 2. Then, in the case of the oblique line shown in FIG. 3, the value becomes zero, and this zero is supplied to the output terminal 20.

For example, in the example shown in FIG.
D) / 2 is [(A + B) / 2- (E + F) / 2] × 1 /
It is understood that the value is the same as 8. As shown in this example, a case where a high-frequency component is predicted from the slopes of low-order coefficients on both sides of the high-frequency component is called interpolation prediction. On the other hand, for example, FIG.
As shown in (B), from the value of (A + B) / 2, [(A +
B) / 2- (C + D) / 2] × 1/4, that is, a high-frequency component is calculated by subtracting a low-order coefficient generated by a value constituting the high-frequency component from another low-order coefficient. The case of making a prediction from the slope is called extrapolation prediction. Here, transfer functions H ₂₁ to H ₂₄ are generated all the signals in the same block. The transfer functions H ₂₁ and H ₂₄ are used for extrapolation prediction. On the other hand, the transfer function H ₂₂ and H ₂₃
Is used for interpolation prediction.

As described above, in the case of the conventional band division filter bank, (C−D) / 2 or (G
−H) / 2, whereas the band split filter bank provides a signal of zero. Therefore, the power of the higher-order coefficient is reduced, and the entropy of the higher-order coefficient is significantly reduced. Moreover, as described later, when decoding, A, B,
C, D, E, F, G, H signals can be decoded. That is, in general, higher order coefficients are coarsely quantized in consideration of human visual characteristics. Therefore, higher order coefficients related to A to H are (AB) / 2, (CD) / 2, (EF) / 2,
There is a high risk of losing (GH) / 2. If this higher-order coefficient is lost, the straight diagonal line shown in FIG. 3 is reproduced as a step-like line shown by a dotted line. Such reproduction appears as block-like or mosaic-like distortion in an image.

FIG. 2 shows an example of the configuration of the band synthesis filter bank in the inverse wavelet transform according to the present invention. The band synthesis filter bank includes a low-order coefficient input terminal 21 for inputting a low-order coefficient, a high-order coefficient input terminal 22 for inputting a high-order coefficient, and a predicted value generation filter for generating a predicted value of a high-order coefficient. 23, an adder 24, a register 25 for delaying low-order coefficients by the delay caused by the predicted value generation filter 23, upsamplers 26 and 27,
Low-pass filter 2 for low-frequency interpolation and extrapolation
8, a high-pass filter 29 and an adder 30 for adding the values of the low-pass filter 28 and the high-pass filter 29
And a signal output terminal 31 for a synthesized signal. Here, the predicted value generation filter 23 constitutes predicted value generation means, the low-pass filter 28 constitutes low-frequency component restoration means, and the high-pass filter 29 constitutes high-frequency component restoration means. Reference numeral 30 denotes a synthesizing unit.

The prediction value generation filter 23 has the same configuration as the prediction value generation filter 16 shown in FIG.
It has to have a transfer function H ₂₁ to H ₂₄ shown in equation (5) to (8). Also, the upsamplers 26 and 27
Is an algorithm reverse to that of the downsamplers 14 and 15 shown in FIG. Furthermore, low-pass filter 2
8 is the same as the low-pass filter 125 shown in FIG.
This is a Haar-based filter having the transfer function H ₃ shown in Expression (3). Also, the high-pass filter 29
This is the same as the high-pass filter 126 shown in FIG. 21, and is a Haar-based filter having the transfer function H ₄ shown in Expression (4).

Next, the operation of the band synthesis filter bank will be described. The low-order coefficient input from the input terminal 21 is supplied to the register 25 and the prediction value generation filter 23. The prediction value generation filter 23 performs filter processing using the transfer functions H _{21 to} H ₂₄ shown in Expressions (5) to (8) in the same manner as the conversion side, and outputs the output from the higher-order coefficient input terminal 22 by the adder 24. Add to the input higher order coefficient. Adder 24
The output of (1) is interpolated by a sample of zero value as appropriate via an upsampler 27, filtered by a high-pass filter 29, and supplied to an adder 30.

On the other hand, the low order coefficient of the output of the register 25 is
Similarly, a sample having a zero value is appropriately interpolated by the upsampler 26 and supplied to the adder 30 via the low-pass filter 28. In the adder 30, the low-order coefficient supplied from the low-pass filter 28 and the high-pass filter 29
Are added, and a signal synthesized with the original signal is output to the signal output terminal 31.

Next, a specific operation of the band synthesis filter bank will be described with reference to the example shown in FIG. Input terminal 21
Include (A + B) / 2 and (C + D) / 2, (A + B) /
2 and (E + F) / 2, (C + D) / 2 and (G + H) / 2
And (E + F) / 2 and (G + H) / 2 signals come in. On the other hand, the input terminal 22 receives a zero signal in the above example. Then, in the predicted value generation filter 23,
[(A + B) / 2- (C + D) / 2] × １ ／, [(A
+ B) / 2- (E + F) / 2] × １ ／, [(C + D)
/ 2- (G + H) / 2] × １ ／, [(E + F) / 2-
A signal of (G + H) / 2] × １ ／ is obtained (provided that the prediction coefficient k is 1 and is not limited by the limiter), and this signal and zero are added by the adder 24.

[(A + B) / 2- (C + D) / 2]
× １ ／ is the same value as (A−B) / 2, [(A + B) /
2− (E + F) / 2] × ８ is the same value as (C−D) / 2, which is the same result as supplying the original higher-order coefficient. Other values are similar. Then, a high-order coefficient is generated by the high-pass filter 29 having the transfer function H ₄ and supplied to the adder 30. That is, (A-
B) / 2, (-A + B) / 2, (CD) / 2, (-C
A value such as + D) / 2 is applied to adder 30.

On the other hand, the low-order coefficient is stored in the register 25 by the delay caused by the predicted value generation filter 23, and after the zero sample is inserted by the upsampler 26, the low-order coefficient is generated by the low-pass filter 28 and added. Input to the container 30. At this time, the low-order coefficients processed by the transfer function H ₃ are (A + B) / 2,
(A + B) / 2, (C + D) / 2, (C + D) / 2,
(E + F) / 2 or the like. Then, as a result of the addition by the adder 30, each value of A, B, C, D, E... Is output from the signal output terminal 31.

Next, an image encoding device and an image decoding device using the wavelet transform device and the inverse wavelet transform device shown in FIGS. 1 and 2 will be described.

As shown in FIG. 4, the configuration of this image coding apparatus is as follows: a horizontal direction conversion unit 41 as a first conversion unit for performing a wavelet conversion, a buffer memory 42 having a two-bank configuration, and a wavelet conversion. A vertical direction conversion unit 43 serving as a second conversion unit, and a quantization unit 44 serving as quantization means
And variable length encoding means 45. Note that a work memory 46 for storing DC coefficients Y ₀ and X ₀ of the immediately preceding block may be added in order to obtain predicted values of high-frequency primary coefficients Y ₁ and X ₁ described later.

In the image coding apparatus, first, an image signal is input to the horizontal direction conversion unit 41 for each 8 × 8 pixel block. As shown in FIG. 5, the configuration of the horizontal direction conversion unit 1 is such that the wavelet conversion device shown in FIG. 1 is connected in multiple stages.

In FIG. 5, an input image signal is first converted into a coefficient 52 representing a low-frequency component and a highest-order coefficient 53 representing a high-frequency component (FIG. _{Y 30, Y 31, Y 32} , represented by Y ₃₃₎ to be frequency division. Low frequency component coefficient 52
Are further converted by the horizontal quadratic wavelet transform device 54 into a horizontal second-order low-frequency coefficient 55 and a high-order second-order coefficient 5
6 (indicated by Y ₂₀ and Y ₂₁ in FIG. 5). Next, the horizontal third-order wavelet transformer 57 converts the low-frequency coefficient 55 to the high-frequency first-order coefficient 58 (Y _{0 in} FIG. 5).
) And a DC coefficient 59 are generated. In this way, the image data of 8 × 8 pixels becomes 8 × 8 coefficient data as shown in FIG. 7 and is frequency-divided in the horizontal direction. The wavelet transform devices 51, 54 and 57 constitute a first wavelet transform device for an image and these devices 51, 5
4, 57, the first wavelet transform step is performed.

A, B, C, D, E, F, G,
In the case of each image signal of H, the coefficient 52 is a = (A
+ B) / 2, b = (C + D) / 2, c = (E + F) /
2, d = (G + H) / 2, each value a, b, c, d
Is equivalent. The highest order coefficient 53 is Y ₃₀ = (A−B) / 2
-Limit ((a-b) / 4), Y 31 = (C-D) /
2-Limit ((ac) / 8), Y ₃₂ = (EF)
/ 2-Limit ((b−d) / 8), Y ₃₃ = (G−
H) / 2-Limit (( c-d) / 4) each value represented by _{Y 30, Y 31, Y 32} , Y 33 corresponds.

The low-frequency coefficient 55 is given by e = (a + b) /
2, the respective values e and f expressed by f = (c + d) / 2 correspond. Further, the high-frequency quadratic coefficient 56 is represented by Y ₂₀ = (ab) /
2-Limit ((e-f ) / 4), Y 21 = (c-d)
/ 2-Limit ((ef) / 4), each value Y
₂₀ and Y ₂₁ correspond. The high-frequency first-order coefficient 58 is represented by Y ₁
= (Ef) / 2, the DC coefficient 59 is Y ₀ = (e + f)
/ 2 respectively correspond. Note that here, Limit
() Is a so-called limiter, which is a function that limits the output to a certain range.

Here, if a work memory 46 is provided to store Y ₀ of the immediately preceding block, the current Y ₀
(= Y _0n ) and the _immediately preceding Y ₀ (= Y _0n-1 ) can be extrapolated, and a limiter can be added as described above. That is, Y ₁ = (ef) /
Can be calculated 2-Limit the Y ₁ as _{((Y 0n-1 -Y 0n} ) / 4) and made.

The image signals A, B, C, D, E, F,
G, H and coefficients 52, 53, 55, 56, 58, 59
Is shown below.

[0074]

(Equation 3)

The coefficients 53, 56, 58 and 59 (eight coefficients in total) are temporarily stored in the buffer memory 42.
Then, it is input to the vertical direction conversion unit 43, and the frequency conversion in the vertical direction is performed by the image second conversion unit having the same configuration as the previous image first conversion unit. The input to the vertical direction conversion unit 43 is performed by raster-scanning the buffer memory 42 in the vertical direction.

As shown in FIG. 6, the vertical direction conversion unit 43 includes a vertical primary wavelet converter 61, a vertical secondary wavelet converter 62, and a vertical tertiary wavelet converter 63 (each of which is a horizontal one).
Next, second and third order wavelet transform devices 51, 54,
57, eight coefficients (one DC coefficient X ₀ , one high-frequency primary coefficient X ₁ , and two high-frequency secondary coefficients X
₂₀ , X ₂₁ , four highest order coefficients X ₃₀ , X ₃₁ , X ₃₂ , X ₃₃ )
Is generated.

The four types of coefficients in the vertical direction are generated for each line in the vertical direction and output to the next quantization section 44. Thus, an image of 8 × 8 pixels is shown in FIG.
The frequency is divided into 16 bands as shown in FIG. Preferably, the buffer memory 42 is composed of two banks. This is because, when data is written to one bank, it is more efficient if the stored contents are extracted from the other bank.

The quantization unit 44 performs quantization by appropriately changing the weight according to each coefficient. In general, low-order coefficients are quantized densely and high-order coefficients are coarsely quantized. The quantized coefficient is input to the variable-length coding unit 45, and a code is assigned according to the magnitude of the appearance probability. Then, it is encoded and output by a variable length code such as a run length code or a Huffman code.

On the other hand, as shown in FIG. 9, the configuration of the image decoding apparatus includes a variable length decoding unit 71, an inverse quantization unit 72 serving as an inverse quantization unit, and a first unit for performing an inverse wavelet transform.
, A buffer unit 74, and a horizontal inverse transform unit 75 as a second inverse transform unit that performs inverse wavelet transform. A work memory 76 for storing the DC coefficients Y ₀ and X ₀ of the immediately preceding block may be provided.

In this image decoding apparatus, first, the variable length coded input data is decoded by the variable length decoding section 71 into quantized coefficients. Subsequently, inverse quantization is performed by an inverse quantization unit 72 to obtain coefficients, and these are decoded by a vertical inverse conversion unit 73 in the vertical direction. Then, the decoded value is stored in the buffer memory 74.

The hardware configuration of the inverse quantization unit 72 is as shown in FIG. That is, a Q-matrix register unit 7 that loads a quantization scale from a host bus and receives a strobe signal serving as a timing signal.
2a and a register 72b for storing a strobe signal,
72b and a register 72c for temporarily storing the quantization coefficient
, A multiplier 72d for multiplying the quantization coefficient from the register 72c by the signal from the Q-matrix register section 72a to obtain an inverse quantization coefficient, and a register 72e for temporarily storing the value.

The vertical inverse transform unit 73 is configured as shown in FIG. 11, and has a configuration in which the wavelet inverse transform device shown in FIG. 2 is connected in multiple stages. In FIG. 11, a DC coefficient X ₀ and a high-frequency first-order coefficient X ₁ are input to a vertical tertiary wavelet inverse transform device 81 and synthesized to obtain a low-frequency coefficient 82. The low frequency coefficient 82 and the high frequency quadratic coefficient 8
X ₂₀ and X _{21, which} are 3, are supplied to the vertical quadratic wavelet inverse transform unit 84 to obtain a low-frequency coefficient 85.

X ₃₀ , X ₃₁ , X ₃₂ , and X _{33, which} are the low-frequency coefficient 85 and the high-frequency tertiary coefficient 86, are supplied to the vertical primary wavelet inverse transform unit 87, and the last vertical inverse transform is performed. Is performed. Then, the frequency synthesis in the vertical direction is completed. It should be noted that the wavelet inverse transform devices 81, 84 and 87 constitute a first wavelet inverse transform device for an image, and these devices 8
1, 84, 87 implement the first inverse wavelet transform step.

FIG. 12 shows a specific hardware configuration in the case where this vertical inverse conversion unit 73 is realized by one device.

The inverse vertical transforming unit 73 receives the inversely quantized strobe signal and controls the respective components.
a, a predicted value generation unit 73b that generates a predicted value with a limiter from the low frequency coefficients 77 and 80, and an addition unit 73c that adds the predicted value with a limiter and the inverse quantization coefficient from the predicted value generation unit 73b. , Calculates the next low-frequency component from the high-frequency component obtained by the addition, and outputs a calculation unit output signal VACMD.
And a register unit 73 for outputting the output signals IVRD0 to IVRD7, collectively storing all the data obtained by the operation, and outputting the output signals IVRD0 to IVRD7.
e and the stored data as 1
And a selector 73f that projects one by one.

When a buffer memory 76 is provided, the signal input to the predicted value generation unit 73b by a dotted line is a signal corresponding to the DC coefficient X ₀ of the immediately preceding block stored in the buffer memory 76. On the other hand, the signal indicated by the dotted line output from the operation unit 73d is a signal corresponding to the DC coefficient X ₀ of the current block used for the next block.

The operation unit 73d comprises eight accumulators 90 to 97
The operation unit input signal VACMIN obtained by the adder 73c is supplied to each of the accumulators 90 to 97. The accumulator 90 includes an AND circuit 90a and an adder 90b.
And a register 90c. The accumulator 91 is
AND circuit 91a, adder / subtractor 91b, register 91
c. The accumulators 92 to 97 correspond to the accumulator 91.
It has the same configuration as. And each accumulator 90-9
7 is an initialization signal VAINIT, which accumulators 90 to 97
Accumulator selection signal VAR for selecting whether to activate
STB is controlled by a subtraction selection signal VASUB for selecting the accumulators 90 to 97 for performing the subtraction.

When the initialization signal VAINIT becomes active, each of the accumulators 90 to 97 discards the past accumulated value and stores the data of the operation unit input signal VACMIN directly in the registers 90c to 97c. Accumulator selection signal VAR
When the STB is activated, the corresponding accumulators 90-9
7 is updated and the summation is executed. By activating the subtraction selection signal VASUB, the addition / subtraction units 91b to 91b of the corresponding accumulators 91 to 97 are set.
97b performs subtraction. Note that the accumulator 90 has an adder 90b and executes only addition.

As described above, in each of the accumulators 90 to 97, the AND circuits 90a to 97a become zero by the initialization signal VAINIT "1" indicating active, and the data of the operation unit input signal VACMIN is sent to the registers 90c to 97c. Store. Thereafter, the initialization signal VAINIT becomes “0”.
The values of the registers 90c to 97c are transmitted to the adder 90b and the adders / subtractors 91b to 97b, and the subtraction selection signal VASU
B subtracts from or adds to the arithmetic unit input signal VACMIN.

Next, the process performed by the vertical inverse converter 73 will be described with reference to the tables shown in FIGS. The number of cycles shown in each table is based on the case where eight inverse quantization coefficients C ₀ , C ₁ , C ₂ , C ₃ , C ₄ , C ₅ , C ₆ , and C ₇ are input every cycle ( The quantized coefficient from the variable-length decoding unit 71 is output at every cycle at the highest speed), and the odd-numbered cycles not shown are wait cycles. Actually, there are many cases where two or more wait cycles are included. Also, in each table, the calculation unit input signal that is obtained by multiplying the inverse quantization coefficient by 8 is used because the calculation is performed at the fixed point, so that the calculation accuracy is maintained by using the first digit. And
Finally, it is reduced to 1/8 and restored.

First, the initialization signal VAINIT is "1"
, The past accumulated values of the accumulators 90 to 97 are discarded.
Also, the accumulator selection signal VARSTB is "11111111".
1 "activates all accumulators 90-97,
When the operation selection signal VARSTB is "00000000", all the adders / subtractors 91b to 97b are in an addition state. In this state, the DC coefficient C ₀ (= X ₀ ) among the inverse quantization coefficients is input. At this time, since the prediction is not performed, the predicted value is “0”, and the operation unit input signal VACMIN becomes D
₀ = 8 × C ₀ . The D ₀ is each register 90c~97
c. Since the operation unit output signal has not yet been generated, there is no signal, and therefore, there is no register output.

In the next cycle 2, the high-order primary coefficient C ₁ (=
X ₁ ) is input. At this time, when there is no work memory 76, the prediction is not performed and the predicted value is “0”. In addition,
When the work memory is present, the steps are the same as those in the following steps, and the description is omitted. Since the predicted value is “0”, the operation unit input signal VACMIN is D ₁ = 8
× a C _1.

On the other hand, the initialization signal VAINIT is "0"
The output of the arithmetic unit output signal VACMD is fed back to the adder 90.
b and adders / subtractors 91b to 97b are input as they are. Further, the accumulator selection signals VARSTB are all active, and the accumulators 90 to 97 are activated.
Further, the subtraction selection signal VASUB is "1111000".
From the top, the three adder / subtractors 91b to 93b perform addition, and the remaining four adder / subtractors 94b to 97b perform subtraction. The adder 90b always performs only addition.

[0094] Due to such a state, the operation unit 73d calculates
From D1 Prefecture as the arithmetic unit input signal and each value D0 of the register 90C～97c, cumulative sum unit 90-93 is Da = D ₀ + D ₁
, And the accumulators 94 to 97 perform subtraction of Db = D ₀ −D ₁ . Here, Da and Db correspond to the low-frequency coefficient 82, and in the horizontal direction, correspond to the low-frequency coefficient 55 (= e, f). Then, the registers 90c to 93c store the value Da.
And the registers 94c to 97c store the value Db.

In the next cycle 4, C ₂ (= X ₂₀ ), which is one of the high-frequency quadratic coefficients 83, is input. Predicted value generator 7
3b selects two appropriate Da and Db from the four output signals Da and Db of the operation unit, and obtains P ₀ = Limit ((Da−Db) / 4) as a prediction value with a limiter. As a result, the two signals are added by the adder 73c, and the arithmetic unit input signal VACMIN becomes D ₂ = 8 × C ₂ + P ₀ . The operation unit output signals Da,
Since Db has already been multiplied by 8, it is not necessary to multiply the predicted value P ₀ by 8.

At this time, the initialization signal VAINIT is "0", and the fed back operation unit output signal VACMD is directly input to the adder 90b and the adders / subtractors 91b to 97b. The accumulator selection signal VARSTB is "0"
0001111 ", the accumulators 90 to 93 are active, the accumulators 94 to 97 are inactive, and the registers 94c to 97c are not updated. The subtraction selection signal VASUB is" 0000110 ", and the addition is performed. Table 9
0b and the adder / subtractor 91b perform addition, and the adder / subtractor 92b, 9
3b performs a subtraction process. Note that the adders / subtractors 94b to 97b
Is added with the addition processing signal “0”.
Since 97 is not active, no addition processing is performed.

Since the operation unit 73d is in such a state, the accumulators 90 and 91 add the operation unit output signal Da and the operation unit input signal D _2, and then the value Dc (= Da + D ₂ )
Is stored in the registers 90c and 91c. On the other hand, the accumulators 92 and 93 store the value of the subtraction processing (Dd = Da−D ₂ ) in the registers 92c and 93c. This is the low frequency coefficient 8
5, Dc corresponds to the low-frequency coefficient a, and Dd corresponds to the low-frequency coefficient b.

Next, another high-frequency quadratic coefficient C ₃ (= X
₂₁ ) Enter. The prediction value generation unit 73b has a register therein, and holds the operation unit output signals Da and Db in the previous step. Then, without using Dc and Dd as new operation unit output signals, P ₁ as a predicted value with a limiter is obtained from Da and Db of the previous operation unit output signal.
= Limit ((Da-De) / 4) is obtained. This value P ₁ is the same value as the predicted value P _{0 described} above.
These two signals are added by an adder 73c, and an arithmetic unit input signal V
As ACMIN, D ₃ = 8 × C ₃ + P ₁ is obtained.

At this time, the initialization signal VAINIT is "0", which is the same as the above. On the other hand, the accumulator selection signal VARSTB is “11110000”, the accumulators 90 to 93 are inactive, and the accumulators 94 to 9 are inactive.
7 becomes active. Also, the subtraction selection signal VASUB
Is "1100000", and only the adders / subtractors 96b and 97b perform the subtraction processing. And adders / subtractors 94b and 95b
Performs the addition process. The adder 90b and the adders / subtractors 91b to 93b do not perform the addition processing because the accumulators 90 to 93 are inactive.

The operation unit 73d is brought into the above-mentioned state by these control signals, and receives the value D ₃ as an operation unit input signal.
, The accumulators 94 and 95 perform an addition process to add the value Db stored earlier, and obtain the value De (= Db + D
₃ ) is stored in the registers 94c and 95c. On the other hand, the accumulators 96 and 97 perform a subtraction process and obtain a value Df (= Db−D
₃ ) is stored in the registers 96c and 97c. The value Dc is stored in the registers 90c and 91c.
The values Dd are stored in c and 93c as they are. Each of the values De and Df corresponds to the low-frequency coefficient 85, and also corresponds to each of the low-frequency coefficients c and d in the horizontal direction described above.

In the next cycle 8, C ₄ (= X ₃₀ ) which is one of the high-frequency third-order coefficients 86 is input. Predicted value generator 7
3b is P ₂ = Limit ((Dc−Dd) / as a predicted value with a limiter from Dc and Dd of the operation unit output signal.
Obtain 4). Then, the adder 73c adds the two signals and obtains a value D ₄ = 8 × C ₄ + as a calculation unit input signal VACMIN.
Get the P _2. At this time, the accumulator selection signal VARSTB is “00000011” and the accumulators 90 and 91
Are active, and the accumulators 92 to 97 are inactive. Further, the subtraction selection signal VASUB is set to “00”.
0000 ", the adder 90b performs the addition process, and the adder / subtractor 91b performs the subtraction process. The adder / subtracters 92b to 97b perform the processing operation because the accumulators 92 to 97 are inactive. Absent.

When the value D ₄ is input to the arithmetic unit 73d as the arithmetic unit input signal VACMIN, the adder 90b
Is, Dg to the register 90c adds the calculation unit output signal Dc retrieved from the value D ₄ and the register 90c = Dc + D
Store ₄ . The adder / subtractor 91b calculates Dh = Dc
Storing a value of -D ₄ to register 91c. This value Dg
Is the value of the image signal A, and the value Dh is the value of the image signal B.

Next, one of the high-frequency third-order coefficients 86, C ₅
When (= X ₃₁ ) is input, the predicted value P ₃ = Limit ((D
c-De / 8) is obtained, and D ₅ =
8 × C ₅ + P ₃ is input to the calculation unit 73d. As a result, Di = Dd + D is stored in the register 92c of the accumulator 92.
₅ is stored, and Dj =
To store the Dd-D _5. The value Di is the value of the image signal C, and the value Dj is the value of the image signal D. Note that the predicted value P ₃
Since the value Dc is not present in the operation unit output signal VACMD, the value Dc of the previous step stored in the register of the predicted value generation unit 73b is used. A similar situation occurs in cycles 12 and 14 described below, and similar processing is performed.

Next, one of the high-frequency third-order coefficients 86, C ₆
(= X ₃₂ ) is input, Dk = De + D ₆ shown in FIG. 15 is stored in the register 94c of the accumulator 94, and Dl = De.
-D ₆ is stored in the register 95c of the formation sum unit 95.
Next, when C ₇ (= X ₃₃ ), which is the last high-frequency third-order coefficient 86, is input, Dm = Df + D shown in FIGS.
₇ is stored in the register 96c of the accumulator 96, and Dn =
Df-D ₇ is stored in the register 97c of the formation sum unit 97. As described above, all the vertical values of the image signal of the predetermined block are stored in the registers 90c to 97c.

Thereafter, the DC coefficient C ₀ (=
X ₀ ) is input. At this time, the values of the registers 90c to 97c are transferred to the register 73e under the control of the control circuit 73a. On the other hand, the operation unit 73d is initialized when the initialization signal VAINIT becomes "1". Thus, each register 90C～97c, first cycle (= 0 cycles) as well as a value D ₀ is stored.

Next, when the high-frequency first-order coefficient C ₁ (= X ₁ ) is input, the register 9 is registered in the same manner as in the cycle 2 described above.
The value Da is stored in the registers 94c to 97c.
Stores the value Db. At this time, the selector 73f receives the inverse conversion output R under the control of the control circuit 73a.
The value Dg is output as MIN. Thereafter, in each cycle, the selector 73f outputs the inverse conversion output RMIN one after another. That is, Dh → Di → Dj → Dk → Dl → Dm
→ Output in order of Dn.

The signal converted by the vertical inverse converter 73 is temporarily stored in the buffer memory 74. The buffer memory 74 is a transposition memory like the buffer memory 42 and has a two-bank configuration. Specifically, one bank is written in the horizontal direction, for example, and the other bank is read in the vertical direction. Then, the read signal is input to the horizontal inverse transform unit 75, and the second inverse transform unit for image that performs the inverse wavelet transform having the same configuration as the first inverse transform unit for the previous image. Performs frequency synthesis in the horizontal direction.

The horizontal inverse converter 75 has the same configuration as that shown in FIG. That is, a horizontal tertiary wavelet inverse transformer is arranged at a portion corresponding to the vertical tertiary wavelet inverse transformer 81, and a horizontal secondary wavelet inverse transformer is arranged at a portion corresponding to the vertical secondary wavelet inverse transformer 84. Are arranged, and the horizontal primary wavelet inverse transform device is arranged in a portion corresponding to the vertical primary wavelet inverse transform device 87. Also, when the horizontal inverse conversion unit 75 is implemented by one device, the configuration is the same as the configuration shown in FIG.

Then, the DC coefficient Y _{0 in} the horizontal direction, the high frequency 1
High order coefficient Y ₁ , high frequency second order coefficient Y ₂₀ , Y ₂₁ , high frequency third order coefficient Y
Frequency synthesis in the horizontal direction is performed from the eight coefficients ₃₀ , Y ₃₁ , Y ₃₂ , and Y ₃₃ , and each image signal is output.

In the image encoding apparatus and the image decoding apparatus as described above, unlike the conventional wavelet transform and inverse transform apparatus, the work memory can be reduced and the cache hit rate can be increased in the case of software implementation. As a result, the execution speed is increased, and the circuit size can be reduced by hardware implementation. Although a run-length code, a Huffman code, an arithmetic code, and the like are used for entropy coding, it is more preferable to use the following predictive run-length for the purpose of improving the processing speed.

That is, when a binary bit string is input,
"0" or "1" is determined as the dominant symbol, and it is predicted that the number of the dominant symbols is n consecutive. When the prediction is successful, either "0" or "1" is output as a code word, and the encoding is completed. If it deviates, either "0" or "1" is output, the sequence of interest is divided, and the signal state of each divided sequence is confirmed and encoded in the same manner as described above. . Then, the same division and prediction are repeatedly encoded until the prediction is achieved or the division reaches a predetermined number of bits.

Further, in addition to the encoding based on such a principle, a further improved prediction run length may be used. That is, the bit string determined by the prediction bit length n is not encoded at once, but is encoded in multiple stages. For this reason, even if the prediction bit length n increases, it is not necessary to increase the capacity of the decode buffer unit. In performing the encoding, an encoding table in which output signals corresponding to input signals are represented in advance may be provided.
By doing so, the encoding speed is improved.

In the image decoding apparatus, the predicted run length and the improved predicted run length described above are as follows.
The decoding may be performed using an inverse algorithm. By doing so, the decoding speed can be increased and the capacity of the decode buffer can be reduced.

In general, in the wavelet transform, unlike the DCT, a slight correlation remains between the coefficients after the transform. In the technique using Japanese Patent Application No. 8-50634 invented earlier by the present inventor, in order to remove this correlation, the entire image is shown in FIG.
As in the case of 2, after dividing into 10 bands, scanning is performed in consideration of the correlation within the coefficient, and conditional entropy coding based on the immediately preceding coefficient value is performed. On the other hand, in the case of performing block division of about 8 × 8 pixels in order to reduce the work memory as in the present invention, it is difficult to scan in consideration of the correlation between coefficients, and it is difficult to remove the correlation, which may cause a reduction in efficiency. There is. However, in the above-described embodiment, the number of band divisions is increased to 16 as shown in FIG. 8 and each coefficient is made as uncorrelated as possible, so that the efficiency does not decrease.

In the improved Haar transform filed by the inventor, the efficiency of the Haar transform is greatly improved by predicting a higher-order coefficient from a lower-order coefficient. In other words, since prediction is performed from both the past and future directions (= interpolation prediction), a large amount of work memory is required. On the other hand, in the wavelet transform device and the inverse transform device according to the above-described embodiment, and the image encoding device and the image decoding device using the devices, the work memory is largely reduced.

In order to reduce the work memory, as described above, in the above-described embodiment, signal processing is performed by dividing into blocks of about 8 × 8 pixels. The size of this block may be a square block of another size such as 16 × 16 pixels other than 8 × 8 pixels, or a block of another shape such as a rectangle. In performing this block processing, prediction across blocks is not extrapolation prediction (= bidirectional prediction) but extrapolation prediction (= one-way prediction) that is prediction only from the past.

If the prediction is made in one direction, there is a case where the predicted value largely deviates. However, the purpose of the prediction is to reduce blockiness at low frequencies, in which case the predicted value is usually not very large. Therefore, by applying an appropriate limiter to the predicted value, distortion due to mis-prediction is removed while maintaining the effect of reducing block distortion.

As shown in FIG. 17 (A), the output of the limiter becomes constant values Z ₁ and −Z ₁ when the input is equal to or more than a predetermined value and equal to or less than the predetermined value. However, as shown in FIG. 17B, when the input exceeds a predetermined value and becomes large or small as shown in FIG.
The output may be close to “0”. In this way, when the picture moves greatly, the prediction itself often does not make sense and is often preferable. That is, when a sudden image change occurs, the predicted value becomes large, but in such a case, it is better to leave the correction value as it is, rather than perform the correction with the predicted value poorly. This is because it is often a value.

When the input signal includes an image color signal, the highest order coefficients Y ₃₀ , Y ₃₁ , and Y in the horizontal and vertical directions are used.
_32, the _{Y 33, X 30, X 31} , X 32, X 33 can be to perform thinning by not transmitting. In this case, on the image decoding device side, interpolation (interpolation) of the color signal can be performed by the inverse wavelet transform processing of the present invention, and special interpolation processing can be omitted.

The above embodiments are only examples of preferred embodiments of the present invention, but the present invention is not limited to these embodiments, and various modifications can be made without departing from the spirit of the present invention. It is. For example, interpolation prediction (bidirectional prediction)
For, the limiter is not necessarily required because the error is reduced. That is, the limiters of the equations (6) and (7) are not necessarily required.

Further, in each of the above-described embodiments, the predicted value of the higher-order coefficient is generated from the lower-order coefficient. However, the function may be incorporated in the extraction means of the higher-order coefficient. For example, the transfer function shown in equation (10) to the high-pass filter for the higher-order coefficients extracted using a filter according to a transfer function H ₅ shown below to a low-pass filter for the low-order coefficient extraction equation (9) H
The same effect can be obtained by using the filter according to ₆ .

[0122]

(Equation 4)

The low-pass filter 12 has a transfer function H ₀ shown in the following equation (11) instead of the Haar basis, and the high-pass filter 13 has a transfer function H ₀ shown in the following equation (12) instead of the Haar basis. it may be used as the transfer function H _1.

[0124]

(Equation 5)

Further, the vertical inverse transform unit 73 shown in FIGS. 12 to 16 and the horizontal inverse transform unit 75 having the same configuration are also applied to a wavelet inverse transform device and an image decoding device which are not divided into blocks. can do.

When the signal to be processed is a color image signal, it is preferable that the present invention is applied to each of the Y component, U component and V component of color. That is, the RGB component is converted to the YUV component by the encoder, the present invention is applied to the YUV component, and the YUV component is converted to the RGB component in the last part of the decoder.

In processing such a color image, writing to and reading from the buffer memory 74 is performed as shown in FIG.
The method is performed as shown in FIG. That is, at the time of writing,
As shown in FIG. 18A, data of a Y block, a U block, and a V block are sequentially written. The increment of the address at this time is regarded as a normal count-up.
On the other hand, in the reading, as shown in FIG. 18B, the data of each block of YUV is multiplexed at a rate of 8-4-4 for the convenience of conversion from YUV to RGB performed as the last processing. Is read. Therefore, the increment of the address becomes the addressing prior to the row address. When writing is performed in bank 0, bank 1 is performing reading. Also,
A similar method is employed for the buffer memory 42.

In each of the above embodiments, extrapolation prediction and interpolation prediction are always used for block division in both the horizontal and vertical directions. However, only block division in the horizontal direction is performed in the horizontal and vertical directions. Extrapolation prediction is used when crossing between blocks in both directions, and interpolation prediction is used otherwise. In vertical block division, only the lowest order coefficient (= Y1) in the horizontal direction is calculated between blocks in the vertical direction. That is, extrapolation prediction may be performed, and the other method may use a simple Haar transform. In the vertical block division, a simple Haar transform may be used.
Further, conversely, a simple Haar transform may be used between blocks in the horizontal direction, and each interpolation and extrapolation prediction of the present invention may be used in the vertical direction.

Further, in the above-described embodiment, the encoding is firstly performed in the horizontal direction and then in the vertical direction, and the decoding is performed in the vertical direction first. Although the inverse conversion in the horizontal direction is performed, the conversion in the vertical direction may be performed first, and then the conversion in the horizontal direction may be performed at the time of encoding. In this case, when performing decoding, first, inverse transform in the horizontal direction is performed, and then inverse transform in the vertical direction is performed.

Further, the decoding step of the present invention may be a program having a processing procedure and recorded on a computer-readable recording medium such as a CD-ROM. In addition, the encoding step may be similarly programmed and recorded on the same type of recording medium. Further, the programmed decoding process and encoding process may be distributed to the user by communication means, or the user may download the program from a host storing the program.

[0131]

As described above, the image coding apparatus and method according to the present invention generate a predicted value of a higher-order coefficient from a lower-order coefficient and subtract the predicted value from the higher-order coefficient. On the other hand, the image decoding apparatus and the image decoding method according to the present invention add the same predicted value to the higher order coefficient at the time of synthesis.
For this reason, it is possible to reduce the power of the higher-order coefficients, and thereby greatly reduce the entropy of the higher-order coefficients, improve the coding efficiency, and in the high-frequency region of the image,
Even if the higher-order coefficients are lost due to quantization or the like, they are compensated by the function of the predicted value, and the reproduction of the encoded signal is improved.

Since the extrapolation prediction and the interpolation prediction are appropriately used to generate a prediction value, the present invention can be applied to encoding and decoding of a divided image instead of the entire image. . When images can be divided and processed in block units, the required work memory becomes smaller,
Software implementations increase the cache hit rate and increase execution speed. In addition, in the case of implementation by hardware, the circuit scale is reduced, and downsizing and low cost can be achieved.

In addition, at the time of extrapolation prediction, a prediction error can be reduced by setting an upper limit and a lower limit to the prediction value, so that an image can be efficiently encoded and, at the time of decoding, a block-like or mosaic-like image can be obtained. Distortion can be greatly reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an embodiment of a wavelet transform device according to the present invention.

FIG. 2 is a block diagram showing an embodiment of an inverse wavelet transform device according to the present invention.

3A and 3B are explanatory diagrams showing the principle of generating a predicted value of a higher-order coefficient according to the present invention, wherein FIG. 3A is a diagram showing interpolation prediction (bidirectional prediction), and FIG. FIG.

FIG. 4 is a block diagram showing a configuration of an image encoding device using the wavelet transform device of the present invention.

FIG. 5 is a block diagram illustrating a configuration of a horizontal direction conversion unit used in the image encoding device of FIG. 4;

6 is a block diagram illustrating a configuration of a vertical direction conversion unit used in the image encoding device of FIG.

FIG. 7 is a diagram for explaining conversion by a horizontal direction conversion unit used in the image encoding device of FIG. 4;

8 is a diagram for explaining conversion in a vertical direction conversion unit used in the image encoding device in FIG. 4;

FIG. 9 is a block diagram showing a configuration of an image decoding device using the inverse wavelet transform device of the present invention.

10 is a block diagram illustrating a configuration of an inverse quantization unit used in the image decoding device in FIG.

11 is a block diagram showing a configuration of a vertical inverse transform unit used in the image decoding device of FIG. 9;

12 is a block diagram illustrating a specific configuration when the vertical inverse transform unit used in the image decoding device in FIG. 9 is configured by one device.

13 is a block diagram illustrating a configuration of a calculation unit used in the vertical inverse conversion unit in FIG. 12;

14 is a diagram for explaining a vertical inverse transform operation process performed by the vertical inverse transform unit of FIG. 12, and is a table showing signals of cycle 0 to cycle 6 (= first four steps).

FIG. 15 is a diagram for explaining a vertical inverse transform operation process by the vertical inverse transform unit of FIG. 12, and is a table showing signals from cycle 8 to cycle 14 (= next four steps).

FIG. 16 is a diagram for explaining a vertical inverse transform operation process by the vertical inverse transform unit of FIG. 12;
FIG. 13 is a table showing signals from a to a cycle 22 (= the next four steps).

FIG. 17 is a diagram for explaining a function of a limiter used in the image encoding device and the decoding device according to the present invention;
(A) shows the case where the output stuck to the upper limit or the lower limit by the limiter, and (B) shows the case where the output signal approaches 0 again from the predetermined value when the input signal increases.

18A and 18B are diagrams for explaining a method of writing and reading data to and from a buffer memory when processing a color image signal. FIG. 18A illustrates addressing at the time of writing.
(B) shows addressing at the time of reading.

FIG. 19 is a block diagram illustrating a conventional orthogonal transform coding method using DCT.

FIG. 20 is a block diagram showing a conventional wavelet transform device.

FIG. 21 is a block diagram illustrating a conventional inverse wavelet transform device.

FIG. 22 is a diagram illustrating image division by a conventional wavelet transform device.

FIG. 23 is an explanatory diagram showing a principle of generating a predicted value of a higher-order coefficient according to the invention filed by the inventor of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Input terminal 12 Low-pass filter (low-frequency component extraction means) 13 High-pass filter (high-frequency component extraction means) 14 Selector 16 Predicted value generation filter (predicted value generation means) 17 Register 18 Subtractor (subtraction means) Reference Signs List 19 low-order coefficient output terminal 20 high-order coefficient output terminal 21 low-order coefficient input terminal 22 high-order coefficient input terminal 23 predicted value generation filter (predicted value generation means) 25 register 26 separator 28 low-pass filter (low-frequency component restoration) Means) 29 high-pass filter (high-frequency component restoration means) 30 adder (synthesis means) 31 signal output terminal 41 horizontal direction conversion section (first conversion section) 42 buffer memory 43 vertical direction conversion section (second conversion section) Unit) 44 Quantizing unit (quantizing unit) 45 Variable length encoding unit 71 Variable length decoding unit 72 Inverse quantizing unit (inverse quantizing unit) 73 Vertical inverse Section (first inverse transform unit) 73a control circuit section 73b prediction value generation unit 73c adding unit 73d register 73e selector 74 buffer memory 75 horizontal inverse transform unit (second inverse transform unit)

Claims (13)

[Claims] 1. An image encoding apparatus comprising: a low-frequency component extracting means for extracting a low-frequency component from an input image signal; and a high-frequency component extracting means for extracting a high-frequency component from the image signal. A predicted value generating means for generating a predicted value for predicting the component; and a subtracting means for subtracting the predicted value from the high frequency component. A new high frequency component is obtained by the subtracting means, and the new high frequency component is output. The predicted value generation means includes:
An image coding apparatus characterized in that the prediction value is generated by appropriately using extrapolation prediction only from the past and interpolation prediction from the past and the future. 2. The image signal is divided into predetermined blocks, the extrapolation prediction is performed when the prediction for generating the prediction value extends between blocks, and the interpolation prediction is performed when the prediction does not extend between blocks. 2. The image coding apparatus according to claim 1, wherein an upper limit and a lower limit are set for the prediction value obtained by extrapolation prediction. 3. An image encoding method comprising: a low-frequency component extracting step of extracting a low-frequency component from an input image signal; and a high-frequency component extracting step of extracting a high-frequency component from the image signal. Providing a predicted value generating step of generating a predicted value for predicting the component, and a subtraction step of subtracting the predicted value from the high frequency component, obtaining a new high frequency component by the subtraction step, outputting the new high frequency component, and In the prediction value generation step, the prediction value is generated by appropriately using extrapolation prediction only from the past and interpolation prediction from the past and the future. 4. The image signal is divided into predetermined blocks, the extrapolation prediction is performed when the prediction for generating the prediction value extends between blocks, and the interpolation prediction is performed when the prediction does not extend between blocks. 4. The image encoding method according to claim 3, wherein the upper limit and the lower limit are set to the prediction value obtained by the extrapolation prediction. 5. A low-frequency component restoring means for inputting and restoring a low-frequency component of an image signal extracted by wavelet transform, and a high-frequency component restoring for inputting and restoring a high-frequency component of an image signal extracted by wavelet transform. Means for generating a predicted value for predicting a high frequency component from the low frequency component, and an adding means for adding the predicted value to the high frequency component to obtain a new high frequency component. A synthesis means for synthesizing the low-frequency component and the new high-frequency component, wherein the prediction value generation means performs the prediction by appropriately using extrapolation prediction only from the past and interpolation prediction from the past and the future. An image decoding device for generating a value. 6. An input signal is divided into predetermined blocks. When the prediction for generating the predicted value extends between blocks, the extrapolation prediction is performed. 6. The image decoding apparatus according to claim 5, wherein an interpolation prediction is performed, and an upper limit and a lower limit are set for the prediction value obtained by the extrapolation prediction. 7. A low-frequency component restoring step of inputting and restoring a low-frequency component of an image signal extracted by wavelet transform, and a high-frequency component restoring of inputting and restoring a high-frequency component of an image signal extracted by wavelet transform And a prediction value generating step of generating a predicted value for predicting a high frequency component from the low frequency component, and an adding step of adding the predicted value to the high frequency component to obtain a new high frequency component. A synthesis step of synthesizing the low-frequency component and the new high-frequency component, wherein the prediction value generation step generates the prediction value by appropriately using extrapolation prediction and interpolation prediction. Image decoding method. 8. When the input signal is divided into predetermined blocks, the extrapolation prediction is performed when the prediction for generating the prediction value extends between blocks, and when the prediction does not extend between blocks, the 8. The image decoding method according to claim 7, wherein interpolation prediction is performed, and an upper limit and a lower limit are set for the prediction value obtained by extrapolation prediction. 9. A first conversion unit that divides an image signal into blocks each including a predetermined number of pixels, and performs a wavelet transform on one of the vertical and horizontal signals of the divided image signal. A buffer memory for temporarily storing the coefficients of the low-frequency component and the high-frequency component extracted by the first conversion unit, and either the vertical or horizontal signal of the divided image signal read from the buffer memory A second transform unit that performs a wavelet transform of the following, a quantizing unit that quantizes the coefficient obtained by the second transform unit,
A variable-length encoding unit that assigns a code to the quantized coefficient in accordance with the magnitude of its appearance probability and performs variable-length encoding, wherein at least one of the first and second conversion units is used. The conversion unit includes a low-frequency component extracting unit that extracts a low-frequency component from an input image signal, a high-frequency component extracting unit that extracts a high-frequency component from the image signal, and generates a prediction value that predicts a high-frequency component. And a subtraction means for subtracting the predicted value from the high-frequency component. The subtraction means obtains a new high-frequency component, outputs the new high-frequency component, and outputs the predicted value. An image coding apparatus characterized in that the prediction value is generated by appropriately using extrapolation prediction only from the first and second interpolation predictions from the past and the future. 10. A variable-length decoding means for dividing an input signal into a variable-length-encoded image signal divided into blocks each including a predetermined number of pixels, and a variable-length decoding means for performing variable-length decoding on the image signal. Inverse quantization means for inversely quantizing the image signal, a first inverse transform unit for performing inverse wavelet transform on the inversely quantized signal and either the vertical direction or the horizontal direction of the image signal, and the inverse transform thereof An image decoding apparatus comprising: a buffer memory that temporarily stores the divided signal; and a second inverse transform unit that performs inverse wavelet transform on a signal in any other direction of the divided image signal read from the buffer memory. At least one of the first and second inverse transform units, a predicted value generation unit that generates a predicted value for predicting a high frequency component of the image signal from a low frequency component of the image signal; An adder for adding the predicted value to the inversely quantized value to obtain a new high-frequency component; and an operation for decoding a value in the target direction of the image signal using the low-frequency component and the new high-frequency component And an image decoding device. 11. The prediction value generation unit performs extrapolation prediction only from the past when the prediction for generating the prediction value extends between blocks, and performs the extrapolation prediction when the prediction does not extend between blocks.
The image decoding apparatus according to claim 10, wherein interpolation prediction is performed from both the future and the past, and an upper limit and a lower limit are set to the prediction value obtained by the extrapolation prediction. 12. The at least one inverse transform unit includes a register unit that stores all signals in the same predetermined block obtained by the arithmetic unit, and a selector that sequentially outputs each signal in the register unit. The image decoding device according to claim 10 or 11, wherein 13. The image decoding apparatus according to claim 10, wherein the prediction value generation unit has a register for storing the low-frequency component output from the operation unit.