JPH0438181B2 - - Google Patents

Description

[Detailed Description of the Invention] [Industrial Application Field] This invention relates to an image signal vector quantizer that blocks a plurality of image signal sequences and quantizes them in a multidimensional signal space. . [Prior Art] First, the principle of vector quantization will be explained very briefly. Now, K input signal sequences are put together into an input vector X = {x ₁ , x ₂ , . . . x _K }.
At this time, the K-dimensional Euclidean signal space R ^K ( X ∈
R ^K ) N representative points (i.e. output vector)
The set of y _i = {y _i1 , y _i2 , ..., y _iK } is expressed as Y = [ y ₁ , y
₂ ,..., yN _] . From the set of output vectors, the vector quantizer determines and searches for the output vector y _i that is the shortest distance from the input vector (resulting in the least distortion) as follows. if d( X , _yi ) _< d( X , _yj ) _forallj X → yi where d( X , yi ) is the distance (distortion) between the input and output vectors. At this time, _the input vector The set Y of output vectors _yi is obtained by clustering using an image signal sequence that becomes a training model (selection of representative points and quantization of each representative point of the training model are repeated until the total sum of distortion is minimized). ). Furthermore, in order to improve the efficiency of vector quantization and the versatility of the set of output vectors, the average value of the vector is separated and vector quantized after being normalized by amplitude. Hereinafter, a conventional vector quantizer will be described below with reference to a specific configuration example. FIG. 5 is a block diagram showing an example of the configuration of a coding section of a conventional image signal vector quantizer. In this configuration example, d( X ,
As an example of the definition of y _i ), the absolute value distortion is shown in equation (1). d _{( _} ^_ _{_ _} input vector, 5 is the average value component and amplitude component of input vector 2, 16 is a normalized input vector register,
17 is a code table address counter, 18 is a code table address, 45 is a normalized output vector code table memory, 20 is a normalized output vector register, 21 is a parallel subtracter, 22 is a parallel absolute value calculator, 23 is an absolute value distortion a calculation circuit (accumulator); 24 is a minimum distortion detection circuit; 13 is a vector quantization index; 25 is an index latch; 46 is an encoder that collectively encodes the index and the average value component and the amplitude component;
7 is the encoder output signal. FIG. 6 shows an example of the block configuration of a decoding section of a conventional image signal vector quantizer. In the figure, 48 is a decoder that decodes a vector quantization index, an average value component, and an amplitude component, 49 is a normalized output vector, and 4
3 is a mean value separation normalization decoding circuit, and 50 is an output vector. Next, the operations shown in FIGS. 5 and 6 will be explained. First, input vector 2, S {S ₁ , S ₂ ,...S _K }, is a block of multiple input signal sequences.
The average value component and the amplitude component 5 are separated by the average value separation and normalization circuit 3, and the normalized input vector,
Form X={x ₁ , x ₂ ,...x _K }. Here, if the average value component is m and the amplitude component is σ, then for example, m=K ^-1 _K 〓 ^j=1 Sj σ=K ^-1 _K 〓 ^j=1 |Sj−m| x _j =(Sj−m )/σ...(2) (j=1, 2,...,K) An operation such as equation (2) is performed. By means of the mean value separation and normalization processing, the input vector 2 approaches a constant distribution in the signal space, so that the efficiency of vector quantization is improved. Normalized input vector 4 is then latched into normalized vector register 16. Next, code table address counter 1
7 is normalized output vector code table memory 4
The normalized output vector y _i is sequentially read from 5 and latched into the normalized output vector register 20. Then, the absolute value distortion calculating circuit 23 calculates the distortion d _i of the normal input vector X and the normalized output vector y _i from the parallel subtracter 21 and the parallel absolute value calculator 22 using equation (3). d _i = d ( X , y _{i )} = _K 〓 _j ⁼ 1 _| Normalized input vector X
Find the minimum value of distortion d _i with. In other words, the minimum strain is
It is obtained using equation (4). d=min id _i (4) When the normalized output vector y _i resulting in the minimum distortion is detected, a strobe signal is sent to the index latch 25, and the code table address 18, which is the address of the vector, is fetched. The encoder 46 collectively encodes the vector quantization index 13 of the normalized output vector y _i resulting in the minimum distortion d, the average value component and the amplitude component 5 of the input vector, and outputs it as an encoder output signal 47. Next, in the decoding section of FIG. 6, the decoder 48
Vector quantization index 13, average value component,
and the amplitude component 5, and obtain the normalized output vector from the normalized output vector code table memory 45.
Read y _i and normalize output vector register 20
latch on. Furthermore, the average value separation normalization decoding circuit 43 uses the average value component and the amplitude component 5 to generate an output vector 50, S '={S' ₁ , S' ₂ ,..., S' _K }
decrypt. That is, S′ _j =σ・y _ij +m ...(4) However, (j=1, 2, ..., K) [Problem to be solved by the invention] The conventional vector quantizer for image signals is as follows. Since the encoding performance depends on the set of normalized output vectors stored in the normalized output vector code table memory, high efficiency and versatility are required for the output vector. Therefore, in order to handle images with significantly different statistical properties, such as images that mainly consist of documents and charts, images that require delicate gradation expression, or images that use different sensor systems, it is necessary to Although generation has become more sophisticated and the capacity of the normalized output vector code table memory has become larger, there has been a problem in that it is difficult to obtain a universal output vector. This invention was made to solve these problems, and it is possible to efficiently and quickly generate a new set of output vectors when the properties of an input image change significantly, and to generate a new set of output vectors. ,
The object of this invention is to obtain a vector quantizer for image signals that eliminates problems such as versatility and normalization output code table memory capacity. [Means for Solving the Problems] A vector quantizer for image signals according to the present invention includes a code table memory whose contents can be dynamically rewritten, a cos conversion section that performs cos conversion on an input vector, an inverse cos conversion unit that performs inverse cos conversion;
cos stored in the code table memory
From the set of transformed output vectors again cos
a vector quantization section that selects the one that causes the minimum distortion for the transformed input vector and determines the index of the output vector;
A new vector is created by averaging the cos-converted input vectors.
A code table updating unit is provided for writing a set of cos-converted output vectors into the code table memory. [Operation] In this invention, in order to set the input vector as an output vector suitable for the image to be quantized, the input vector is average-separated and normalized.
Perform clustering with cos transformation and reverse
A set of output vectors for mean value separation normalized vector quantization is generated by cos transformation, and this set of output vectors and the result of vector quantization encoding are collectively used as an encoded output. [Example] Hereinafter, an example of the present invention will be described with reference to the drawings. In FIGS. 1 to 4, the same parts as in FIGS. 5 to 6 are indicated by the same reference numerals. In FIGS. 1 to 4, FIG. This is a configuration example.
In the figure, 1 is an image memory in which images to be quantized are stored, 6 is a vector quantizer, and 7 is a
cos conversion unit, 8 is a normalized input vector that has been cos converted, 9 is a code table update unit, 10 is the latest set of normalized output vectors that have been cos converted, 1
1 is the inverse cos transform unit, 12 is the latest normalized output vector, 13 is the vector quantization index, 14
is an encoder, and 15 is an output signal from the encoder. Further, FIG. 2 shows an example of the configuration of the vector quantization section 6. In the figure, 19 is a dynamic normalized output vector code table (first and third code table memories). Further, FIG. 3 shows an example of the configuration of the code table updating section 9. As shown in FIG. In the figure, 2
6 is a register, 27 is a code table address counter, 28 is a code table address, 29 is a dynamic code table (second code table memory), 30 is a register, 31 is a parallel subtracter, 32 is a parallel absolute value calculator, 33 is an absolute value distortion calculation circuit (accumulator), 34 is a minimum distortion detection circuit, 35 is an index latch, 36 is a vector quantization index, and 37 is a vector quantized index for a normalized and cos-converted input vector. 38 is an index counter that counts the number of vectors that have been quantized for each index; 39 is a vector accumulator that calculates the contents of the vector accumulator 37 with the contents of the index counter 38; A division operator 40 calculates the arithmetic mean of the vectors quantized to each index by dividing, and 40 is the average vector calculated in the division operator 39, that is, the average vector updated in the normalized and cos-transformed domain. This is an output vector register that stores a set of output vectors. Further, FIG. 4 shows an example of the configuration of the decoding section in the vector quantizer for image signals according to the present invention. In the figure, 41 is a decoder that decodes the vector quantization index 13, the average value component and amplitude component 5, and the updated normalized output vector set 12 from the encoder output signal 15, and 42 is a normalized output vector. , 43 is a mean value separation and normalization restoration circuit that performs the inverse processing of the mean value separation and normalization circuit 3, and 44 is a reproduced output vector. Next, the operation will be explained. Basically, when images with different properties are input and it becomes necessary to update the set of output vectors, clustering is performed using the input vector group as a model to obtain a new set of output vectors, and that set is used to perform clustering. This method attempts to cope with large changes in the input image by vector quantizing the input image and using vector quantization information and an updated set of output vectors as the encoded output. By using a large model for clustering and repeatedly converging it, noise components and distortions contained in the model are removed, but in order to converge quickly with a relatively small model, clustering is performed after applying cos transformation. . Here I will briefly explain cos conversion.
When considering the signal sequence f(j), (j=0, 1,...M-1), the one-dimensional cos transformation F(u) of f(j) is as follows.
It can be obtained from equation (5). F(u)=2c(u)/M _M-1 〓 ^j=0 f(j)cos[(2j+1)u/2Mπ] ...(5) (u=0,1,...,M-1) However , c(u)=1/√2(u=0) 1(u=1,2,...,M-1) In addition, the inverse cos transform f(j) of F(u) is given by the following equation (6) It can be obtained with f(j)= _M-1 〓 〓 ^u=0 c(u)F(u)cos[(2j+1)u/2Mπ]...(6
) (j=0,1,...,M-1) [For cos conversion, for example, PRATT
“DIGITAL IMAGE PROCESSING” (WILEY
Since the cos-transformed signal sequence is converted from a vector on the spatial axis to a vector on the time axis, noise components superimposed on the high-frequency components, which have low power but are visually important, are removed. They are quickly removed during the clustering process.
It is also effective to round the coefficients of high frequency components by bit slicing at the stage of cos conversion. The set of output vectors obtained by clustering is converted back into a set of normalized output vectors by inverse cos transformation. The above operation will be explained with reference to the drawings. In FIG. 1, the image signal stored in the image memory 1 is read out in the form of a K-dimensional input vector 2, S. In the average value separation normalization circuit 3, the normalized input vector 4 X , the average value component (m), and the amplitude component (σ) are
form 5. Normalized input vector 4 X is vector quantization unit 6
is latched into the normalized input vector register 16 at . Code table address counter 17
is dynamically normalized output vector code table 19 by sequential code table address 18.
The latest normalized output vector y _i is read from and latched into the normalized output vector register 20. The absolute value distortion calculation circuit 23 uses the parallel subtracter 21 and the parallel absolute value calculator 22 to calculate the absolute value distortion d _i using equation (8). d _{i = K} 〓 _j ^{= 1} _| Incorporate into.
The contents of the index latch 25 at the time when the code table address counter 17 has completed one rotation become the vector quantization index 13. _On ^the other hand _{, the} ^{normalized input vector} 4 The above vector X ^c is then processed by the code table update unit 9,
As shown by the similarity between FIGS. 2 and 3, vector quantization is performed in the cos-transformed state, and a vector quantization index 36 is determined. Vector
^c is accumulated in a vector accumulator 37 for each vector quantization index 36.
At the same time, the index counter 38 counts the number of vectors X ^c mapped to each index, and the division calculator 39 calculates the arithmetic mean of the vectors X ^c mapped for each index. This average vector is output vector register 4
The latest set of normalized output vectors 10, y ^c _i , which are held at 0 and have been cos-transformed, are written to the dynamic code table 29 . The initial contents of the dynamic code table 29 are picked up from ^the cos-converted normalized input vector 8Xc . The above process is clustering, and is expressed by equation (9). y ^c _i ⁼ _Σ ( X ^{c )} _{i /} ‖ _ ^_ _{_} _ ^_ _{_} _ ^_ _{_} ( X ^c ) _i represents the vector X ^c quantized to index i, and | X ^c || _i represents the number of vectors X ^c quantized to index i.
N is the number of output vectors. Clustering requires repeating the above process for many input vector groups, but because of cos transformation, a relatively small number of input vectors can be used, and distortion can be avoided without having to repeat many input vectors. and a set of output vectors with less noise. [ y ^c _i ] is inversely cos-transformed in the inverse-cos transformer 11 and written into the dynamic normalized output vector code table 19 as the latest normalized output vector set 12. At this time, the amplitude is renormalized. In the encoder 14, when the vector quantization index 13, the average value component and the amplitude component 5, and the code table are updated, the latest normalized output vector set 12 is also encoded all together, and is outputted as the encoder output signal 15. Output. The decoder 41 first decodes the vector quantization index 13, the average value component and the amplitude component 5, and, if the code table has been updated, the latest normalized output vector set 12. Then, the latest normalized output vector set 12 is written into the dynamic normalized output vector code table 19, and the normalized output vector y _i 42 giving the minimum distortion according to the vector quantization index 13 is the normalized vector output. Latched into register 20. The average value separation normalization restoration circuit 43 uses the average value component m and the amplitude component σ as shown in equation (10) to calculate the output vector 44S'=
Regenerate {S′ ₁ , S′ ₂ , ..., S′ _K }. S' _j =σ・y _ij +m ...(10) (j=1, 2, ..., K) [Effect of the invention] As described above, according to this invention, images with very different statistical properties can be input. Even when given as It is possible to obtain a set of output vectors with less noise through high-speed clustering using a relatively small clustering model without increasing the memory content, and there are effects such as being able to handle various images.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an example of the configuration of the encoding section of the vector quantizer for image signals according to the present invention, FIG. 2 is a block diagram showing an example of the configuration of the vector quantization section, and FIG. 3 is the code table thereof. FIG. 4 is a block diagram showing an example of the configuration of the updating section, FIG. 4 is a block diagram showing an example of the configuration of the decoding section of the image signal vector quantizer according to the present invention, and FIG. 5 is a block diagram showing the configuration of the conventional image signal vector quantizer. FIG. 6 is a block diagram showing an example of the structure of an encoding section. FIG. 6 is a block diagram showing an example of the structure of a decoding section of a conventional image signal vector quantizer. In the figure, 1 is an image memory, 2 is an input vector, 3 is an average value separation and normalization circuit, 4 is a normalized input vector, 5 is an average value component and an amplitude component, 6 is a vector quantization unit, and 7 is a cos conversion unit , 8 is a normalized input vector that has been cos-converted, 9 is a code table update unit, 10 is the latest normalized output vector set that has been cos-converted, 11 is an inverse cos transform unit, and 12 is the latest normalized output vector set. , 13 is a vector quantization index, 14 is an encoder, 15 is an encoder output signal, 16 is a normalized input vector register, 17 is a code table address counter, 18 is a code table address, 19 is a dynamic normalized output vector code table (first and third code table memories), 20 is a normalized output vector register, 21 is a parallel subtracter, 22
is a parallel absolute value calculator, 23 is an absolute value distortion calculation circuit,
24 is a minimum distortion detection circuit, 25 is an index latch, 26 is a register, 27 is a code table address counter, 28 is a code table address,
29 is a dynamic code table (second code table memory), 30 is a register, 31 is a parallel subtracter, 32 is a parallel absolute value calculator, 33 is an absolute value distortion calculation circuit, 34 is a minimum distortion detection circuit, and 35 is a Index latch, 36 is vector quantization index, 37 is vector accumulator, 38
is an index counter, 39 is a division operator,
40 is an output vector register, 41 is a decoder,
42 is a normalized output vector, 43 is a mean value separation normalization and restoration circuit, 44 is an output vector, 45 is a normalized output vector code table, 46 is an encoder, 47 is an encoder output signal, 48 is a decoder ,
49 is a normalized output vector, and 50 is an output vector.

Claims (1)

[Claims] 1. An image memory that can store an image signal sequence to be encoded, block it into multiple blocks and read it out as an input vector, and separate the average value from the input vector, normalize it by amplitude, and create a normalized input vector. and a first code table memory capable of dynamically rewriting the contents of the normalized output vector after the mean value separation, and the first code table memory stores the contents of the normalized output vector after the mean value separation. Selecting from a set of normalized output vectors after mean value separation, the normalized output vector after mean value separation that causes the minimum distortion with respect to the normalized input vector processed by the mean value separation and normalization circuit; a vector quantization unit that outputs a vector quantization index of the normalized output vector after the mean value separation; a cos conversion unit that performs cos conversion on the normalized input vector after the mean value separation; and a cos conversion unit that converts the normalized input vector after the mean value separation. a second code table memory capable of dynamically rewriting the contents of the input vector;
From among the set of cos-transformed normalized output vectors, the index of the average-separated normalized output vector after cos-transformation that produces the minimum distortion with respect to the cos-transformed normalized input vector processed by the cos-transformed unit is determined. determined and mapped to the same index.
a code table update unit that performs vector addition and averaging of the cos-converted normalized input vectors and writes the result to the second code table memory as a set of the latest cos-converted normalized output vectors; The latest set of normalized output vectors that have been cos-transformed thus obtained is inversely cos-transformed, and the obtained set of latest normalized output vectors is supplied to the first code table memory of the vector quantization section. an encoder that inputs an inverse cos transform unit, a vector quantization index obtained from the vector quantization unit, the latest set of normalized output vectors, and the average value component and the amplitude component, and encodes and transmits the input;
a decoder for decoding the encoder output signal sent from the encoder; and a third code table memory capable of writing the latest decoded normalized output vector set using a vector quantization index. and an average value for reproducing an output vector using the normalized output vector read from the third code table according to the decoded vector quantization index and the decoded average value component and amplitude component. A vector quantizer for image signals, comprising a separation normalization and restoration circuit.