JPS63116585A - Code table producing system - Google Patents

Abstract

PURPOSE:To reproduce a decoding picture signal by reversely transforming each vector in a transformation area included in a code table for quantizing vectors and mapping the vector in the spatial area included in the code table. CONSTITUTION:Each vector included in the code table in a transformation area for quantizing vectors with respect to the transformation coefficient at every divided band group is subjected to inverse orthogonal transformation, and transformed into the code table consisting of vectors in the spatial area. Since an index signal transmitted from a transmission side is generated in the spatial area for the unit of bands, a reception side can decode a picture signal in the spatial area, which corresponds to each band group, only if inverse vector quantizers 71-73 inversely transform the vector. An adder 74 adds and synthesizes the decoded picture signals in each band group. Thus the decoded picture signal including signals in all band groups can be reproduced.

Description

[Detailed description of the invention] (Industrial application field) The present invention relates to encoding of moving image signals.

(Prior Art) Conventional vector quantization uses the amplitude pattern within a block of input pixels and the statistical properties of the amplitude pattern to generate a signal contained in a code table prepared in advance. It was executed using pattern matching with patterns. In other words, it was pattern processing in the spatial domain. In such vector quantization, Yoshibe Yamada et al.'s paper "Vector quantizer design for image signals" Journal of the Institute of Electronics and Communication Engineers 83/8 Vol. j66-B
The method described in No. 8 P2O3-972, etc. can be used.

Separately, the image signal is first subjected to orthogonal transformation to obtain an image signal in the transform domain, vector quantization is performed on this, a quantization index is transmitted, and the receiving side reproduces the vector in the transform domain from this index. Furthermore, there is also a method that combines orthogonal transformation and vector quantization to obtain the original image signal by performing orthogonal inverse transformation. An example of this is the paper "Vector Quantization of Images Using Discrete Cosine Transformation" by Kiyoharu Aizawa et al., Institute of Electronics and Communication Engineers, Image Engineering Research Group Research Report IE84-8f;, pages 65-71.

(Problems to be Solved by the Invention) In conventional B/L quantization, amplitude patterns within a block consisting of a plurality of input pixels and various amplitude patterns stored in a code table prepared in advance are used. Vector quantization was performed by pattern matching. When performing vector quantization, if compression encoding is performed to less than 0.5 bits per sample, image quality will deteriorate significantly, so information of at least 0.5 bits or more is usually allocated. In this case, if vector quantization is to be performed on a block of 8 samples x 8 samples, for example, a very large code table of 32 bits will be required. In reality, if the Read QnlyMemory (RO
M) can be used, and even if the number of output bits is sufficient, there will be 2 to the 22nd power, or approximately 4 million ROs.
M is required, and even if a code table could be realized theoretically, it would be enormous in size from a hardware perspective, making it unrealistic (there were three practical reasons, especially for vector quantization in the spatial domain). It was unrealistic.

In the vector quantization in the transform domain described above, frequency components are grouped little by little to form a band, and the vector I and
It is also possible to fit the code into a code table that can be realized by applying vector quantization, but in this case, hardware for orthogonal transformation and vector quantization is required. Both are realized with the same hardware scale, but since both are required, they are quite large. This conventional example will be explained next.

As shown in Fig. 3, the input image signal or the image signal such as the differential signal obtained by using the correlation between screens (predictive coding that uses the correlation between screens) is divided into each block of a predetermined size. Perform orthogonal transformation. The orthogonally transformed signal is
As shown in Figure A, the upper left of the block is the DC component, followed by the low frequency component, and as you move toward the lower right, the middle and high frequency components.

The signals of 92 professionals, which are composed of signals from DC components to high-frequency components, are divided into band groups (bands) determined as shown in Figure 2B.For example, 8 samples If the signal of a block of ×8 samples is divided into three band groups, each band group will have about 20 samples, and if each sample is given an average of 0.5 bits of information, then by preparing a 10-bit code table, , vector quantization can be performed.In the case of low bit rate transmission, the high-frequency components in the lower right portion are often not encoded from the beginning.

Return to Figure 3. The orthogonally transformed signal is divided into, for example, three band groups of 0.1.2. The orthogonal transform coefficients divided into band groups are vector quantized by a vector quantizer for each band, and an index signal for each band is output and sent to the receiving side. On the receiving side, the index signals of each band sent from the transmitting side are subjected to inverse vector quantization to reproduce the band-divided orthogonal transform coefficients for each band group. The orthogonal transform coefficients reproduced by the inverse vector amount → converter for each band are added and combined by an adder, and become the orthogonal transform coefficients before band division, with vector quantization distortion added. The orthogonal transform coefficients synthesized by the adder are sent to an inverse orthogonal transformer and are inversely transformed to reproduce a decoded image signal in the spatial domain.

(Means for Solving the Problems) According to the present invention, when encoding an image signal or an image signal with reduced redundancy, mapping to a transform domain using any transform represented by orthogonal transform, Divide multiple components in the transform domain into multiple groups, perform clustering on each group to obtain a vector quantization code table, and calculate each vector in the transform domain included in this code table. A code table creation method is obtained in which the inverse operation of the conversion is performed on the code table to map the code table to a code table in a spatial domain.

(Function) The present invention further advances the conventional idea of combining orthogonal transformation and vector quantization. In other words, a vector quantization code table that takes orthogonal transformation and division into band groups into consideration is created in advance by the process shown in Figure 1, and then a vector quantizer for each band as shown in Figure 4 is used. This is an attempt to perform band-divided vector quantization. Regarding this principle, the first
This will be explained using figures. First, we will explain how to create a code table. First, the image signal is orthogonally transformed, the transform coefficients are divided into N predetermined hands, and a code table in the transform domain is created for vector quantization of the transform coefficients for each band group. A so-called LBG algorithm is used to create the code table.

The obtained code table is then subjected to inverse orthogonal transformation for each vector contained therein to convert it into a code table consisting of vectors in the spatial domain. For example, if the code table in the conversion area is 10 bits, then 1
A spatial domain code table with 0.024 types of vectors is obtained.

Corresponding to each band group created in this way ('?
) According to a vector quantizer that uses a code table, as shown in FIG. 4, for an image signal in the spatial domain,
- It is possible to easily perform the combination of orthogonal transformation and vector quantization shown in FIG. 3 by simply using ordinary vector quantizers in parallel. Also, on the receiving side, since the index signal for each band sent from the transmitting side is generated in the spatial domain, the image signal in the spatial domain corresponding to each band group can be obtained by simply inversely transforming it with an inverse vector quantizer. can be decrypted. Then, by adding and combining the decoded image signals of each band group using an adder, a decoded image signal containing the signals of all band groups, that is, the original image signal is obtained except for some distortion due to vector quantization. be able to. This principle will be briefly explained using FIG. 4. Image signal human power f
consists of components of only three groups, band groups O11 and O2, and from each band group, the quantization index K
It is assumed that either + or F2 is output to n, respectively.

Also, C? in the spatial domain for each index rK'o, K1, F2? ) vectors B, Ashi1, and A2 correspond to each other. Yo + ¥? ++L=F, and ν is a vector formed by pixels within a block when the image signal f is divided into blocks.

Q(・) is a vector quantization operator, Q-1(・)
Letting be the inverse operator, then holds true.

Also, the orthogonal transformation operator is T and its transposed form is 'I't.

When the inverse operator is T-1, the two-dimensional orthogonal transformation below is defined as J=T-F-Tt (2). However, J is a vector in the transform domain, and its components are usually as shown in FIG. 2A, taking frequency as an example. Therefore, Ja, , B, and J2 are arbitrary bits of each band group. Of course, these sums are 0+Ryo and 1+Ryo2.

Also J mouth=T-FO・7’t J1=T-Fl・Tt (3) J2=T-F2・Tt holds true.

A method of finding an optimal index in the sense of minimizing error power in vector quantization in band group 0 from input signal vector F will be described below. Generally, if a vector in band group O is expressed as Ko, the search for the optimal index at this time is as follows.

When the error power is P, I → P= (Q-1(KO) -F)" = l Q-'(K'o) -FO-Fl-F212
= cQ-1 (Kg: + -"i' mouth) 2-2
(Q-1(Kff)-lower (1)・CFl+F2'1+
(F1+F2)" (4), and in the second term of the final equation, both q-1(Kn) and Fo are components of band group O.

Since the unique components obtained by orthogonal transformation are mutually independent, there is no correlation between different components. Therefore, there is no correlation between different band groups.

This result (Q-1(Kn)] ・ (F, +F2) = 0 [
F port] ・ CFI + F2) = 0, and equation (4) becomes P = (Ql (Kn) -?o:] 2+ ('F
+172) becomes 2. This is because from the correspondence between the code tables in the transform domain and the spatial domain, Q-' (KN) = T'S, and considering in the transform domain, Tt = T-1 in the Hermitian matrix, [F1] ・CF+:l = (T-1・J'o・T)
・[T-1・Jl・T] −T−1・J′o−Jl・T =0.

As a result, the input image signal F is divided into Fo, Fl, F2 for each band.
There is no need to divide the band groups in advance and perform vector quantization for each band group.If you perform vector quantization in parallel for each band group, indexes Ko and K1 that minimize the error power can be obtained. , F2
are obtained simultaneously and independently.

Therefore, using this method, it is possible to realize a system that combines orthogonal transformation and vector quantization by simply performing vector quantization in the spatial domain and its inverse quantization, and as a result, the hardware can be made smaller. Very effective.

Furthermore, in encoding using conventional vector quantization, when the motion of the moving image becomes rapid and the amount of information generated becomes large, it becomes impossible to match the speed of the transmission path.
In some cases, encoding is unavoidably stopped, resulting in discontinuity in the encoded image and unnatural motion, which is very annoying to the eyes.

In the present invention, for example, the moving area of a moving image is calculated for each screen, and if the moving area of the moving image becomes large and excessive information is likely to be generated, the resolution of the moving part may be slightly reduced. Utilizing the human visual characteristic of not becoming a vector quantum Cl2), it is possible to stop converting only the vector quantum Cl2) to a band group of high frequency components or high and middle frequency components. As a result, the generated information can be smoothly controlled, and sudden generation of information can be suppressed, making it possible to perform smooth encoding of motion.

The motion area detection method for determining the motion area is as follows:
Some are already known, for example, patent application 1984-1.
As described in the specification of No. 94110, "Moving image signal motion/static separation device," the screen is divided into blocks of a predetermined size, and the absolute values of the frame difference values of each pixel in the block are added within the block. There is a method of determining the block by comparing the result of this addition with a predetermined threshold value.

Alternatively, a method called a gradient method can be used to obtain a pixel-by-pixel motion vector from the luminance gradient within a frame and a frame difference value, and a set of pixels for which this motion vector is not zero can be used as a motion region. Any method may be used for motion detection in the present invention.

(Embodiment) An embodiment of the present invention will be described in detail with reference to FIGS. 5, 6, 7, and 8.

An input moving image signal is supplied to the scan converter 1 via line 100 in FIG. A scan converter 1 scan-converts an input moving image signal into blocks of a predetermined size and supplies them to a motion detector 2 and a delay 4. The motion detector 2 is connected to the scan converter 1
The absolute value of the frame difference value of each pixel in the block of the block-formed moving image signal supplied from the block is added within the block, and a motion block is detected by comparing the magnitude of this result with a predetermined threshold value. The method of motion block detection is not limited to this method.

A signal indicating a motion block output from the motion detector 2 is supplied to a control circuit 3. The control circuit 3 adds up the number of motion blocks of the screen to be encoded, and calculates the motion area of this screen. A signal indicating the width of the band limit by comparing the calculated value with a predetermined threshold value is connected to the line 36.
For example, as shown in FIG. When the movement area is 20% or more, a signal indicating the width of the band limit can be determined. The blocked moving image signal supplied from the subtracter 5 is delayed by an amount corresponding to the time required for motion detection and determining the width of the band limit, and is then supplied to the subtracter 5.

The subtracter 5 subtracts the human-powered moving image signal supplied from the delay 4 and the prediction signal supplied from the frame memory 9 to obtain a prediction error signal. The prediction error signal at the output of the subtractor 5 is supplied to the vector quantizer 6 via line 56. As shown in FIG. 6, the vector quantizer 6 is assumed to be divided into three band groups, eg, low band, middle band, and high band, and perform vector quantization for each band. The method of dividing this band loop is not limited to that shown in FIG. 2B. The vector quantizer 6 includes, for example, 3 as shown in FIG.
It can be configured with two vector quantizers. For example, the vector quantizer 61 is used for low frequencies, the vector quantizer 62 is used for middle frequencies, and the vector quantizer 63 is used for high frequencies. The code table for each vector I/L quantizer is created by extracting each band's signal from the orthogonally transformed signal, creating a vector quantization code table in the transform domain for each band, and then A code table is created in advance, consisting of vectors obtained by inverse orthogonal transformation of each vector and converted into a spatial domain. In this way, vector quantization can be performed by directly dividing the spatial domain moving image signal into bands using only the vector quantizer in the spatial domain. The prediction error signal supplied from the subtractor 5 via line 56 is supplied to a vector quantizer 61, 62, 63. The vector quantizer 61 performs vector quantization on the low frequency component signal included in the spatial prediction error signal supplied from the subtracter 5. The vector quantizer 62 performs vector quantization on mid-range components included in the prediction error signal. The vector quantizer 63 performs vector quantization on high frequency components included in the prediction error signal. As mentioned above, components in different band groups do not interfere with detection in other bands.The index signal of each band, which is the output of each vector quantizer, is .672.67
3 to an inverse vector quantizer 7 and a variable length encoder 10. Further, the vector quantizer 7 is connected to the control circuit 3.
Bandwidth limiting is performed in accordance with a signal indicating the bandlimiting width supplied via line 36 from . For example, if the signal indicating the width of the band limit supplied by the line 36 indicates that the band limit is to be performed on a high frequency signal, the output of the vector quantizer 63 is set to zero or zero. By outputting an index representing , high-frequency signals can be deleted. In addition, if the signal indicating the width of the band limit indicates that the band limit of the high-frequency and mid-range signals is to be performed, the output of the vector quantizer 62 and 63 is set to zero or represents zero. By outputting the index, high-frequency and middle-frequency signals can be deleted and the amount of information can be reduced. Bandwidth limiting can be performed in this way.

The inverse vector quantizer 7 includes three inverse vector quantizers 71, 72, and 73, and an adder 74, as shown in FIG.

Each inverse vector quantizer is a vector quantizer 6
The vector quantization inverse transform is performed on the index signals for each band supplied from the adder 74, and the signals divided into bands in the spatial domain are supplied to the adder 74, respectively. Adder 7
4 adds the signals for each spatial domain band supplied from the inverse vector quantizers 71, 72, and 73 to reproduce the original prediction error signal. The output of adder 74 is supplied to adder 8 via line 78 as the output of inverse vector quantizer 7 .

Return to Figure 5. Adder 8 adds the signal supplied from inverse vector quantizer 7 and the predicted signal supplied from frame memory 9 to generate a locally decoded signal. The locally decoded signal at the output of adder 8 is supplied to frame memory 8 . The frame memory 8 delays the signal supplied from the adder 8 by about one frame time.The subtracter 5 and the adder 8
supply to.

Next, the variable length encoder 10 encodes the index signals of each band supplied from the vector quantizer 6 using efficient codes such as Huffman codes, and supplies the encoded signals to the buffer memory 11 . The buffer memory 11 is
The variable length code supplied from the variable length encoder 10 is output to the transmission line via the line 110 while matching the speed of the transmission line.

The number of vector quantizers for each band group is 2.
It does not matter how many you use as long as it is more than 1.

That is, there is no particular limit to the number of divisions into band groups.

(Effects of the invention) As explained in detail above, if vector quantization is performed directly on a block of 8 samples x 8 samples, for example, a code table of at least 30 bits or more will be required, resulting in an enormous amount of hardware. It was so unrealistic. However, if the present invention is used to divide a spatial domain signal into multiple bands and perform vector quantization for each band, the code table can be realized with at least 10 bits or less, which is very practical. It is. Also, in terms of hardware size, vector quantization with band division can be achieved using only a vector quantizer and an inverse vector quantizer, which is very effective when considering miniaturization. Furthermore, although vector quantization is performed in the spatial domain, execution/non-execution can be controlled for each band group in the transformation domain, making it easy to control the amount of generated information and image quality. When the present invention is put to practical use in this way, the effects are extremely large.

[Brief explanation of the drawing]

1, 2A and 2B, 3 and 4 are detailed illustrations of the present invention, and FIGS. 5, 6, 7, and 8 are
FIG. 2 is a diagram illustrating the present invention in detail. In the figure, 1...Scan converter 2...Motion detector 3...Control circuit 4...Delay 5...Subtractor 6...Vector quantizer 7...
Inverse vector quantizer 8...Adder 9...Frame memory 10...Variable length encoder 11...Buffer memory 61...Vector quantizer 62...Vector quantizer 63... - Vector quantizer 71... Inverse vector quantizer 72... Inverse vector quantizer 73... Inverse vector quantizer 74... Adder,
It is. ζ21) Figure 6 Figure 7

Claims (1)

[Claims] When encoding an image signal or an image signal with reduced redundancy, it is mapped to a transform domain using an arbitrary transform such as orthogonal transform, and multiple components in the transform domain are combined into multiple groups. The code table for vector quantization is obtained by performing clustering on each group, and the inverse operation of the above transformation is performed on each vector in the transform domain included in this code table to create a space for the code table. A code table creation method characterized by mapping to a code table in a region.