JPH0830585A - Methods for encoding and decoding picture signal by orthogonal base matrix calculating method - Google Patents

Abstract

PURPOSE:To provide picture signal encoding and decoding methods capable of obtaining sufficiently large encoding efficiency and suppressing the generation of a specific noise so called mosquito noise even when a picture signal has not a steady feature but a local one and to provide a normalized orthogonal base matrix calculating method to be used for the orthgonal transformation of the encoding and decoding methods. CONSTITUTION:In a method for successively finding out N N-order base row vectors b(1) to b(N) which are mutually orthogonal in an N-order square normalized orthogonal base matrix B so that the n-th base row vector b (n) maximizes b (n)Rb in an N-order square covariant matrix R and intersects with all other vectors b(1) to b(n-1), each different N-order square covariant matrix R(n) is used. Orthogonal transformation based upon the obtained base matrix B is applied to a picture element block by an orthogonal transformation means 302.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image signal encoding method and a decoding method using an orthonormal basis matrix calculating method, which is used in a videophone, a videoconference and the like.

[0002]

2. Description of the Related Art There has been a strong demand for efficient encoding and decoding for transmitting and storing digital image signals. In recent years, a method using orthogonal transformation has become common as encoding and decoding of image signals. In particular,
Discrete cosine transform (DC
T) is often used.

In the orthogonal transformation, an N-order square orthonormal matrix A called a basis is used to perform an operation APA ^{t on} an N-order square block P cut out from a certain two-dimensional signal,
This is an operation for obtaining the Nth-order square coefficient matrix C. The DCT, which is a kind of orthogonal transform, is represented as a special form of the Karhunen-Reeve transform (KLT) which is said to give the best coding efficiency to a signal sequence having a certain correlation. The base row vector k (n) of the base matrix K used for the KLT gives the maximum value of the operation k (n) Rk ^t (n) for the covariance matrix R of the signal sequence according to the stationary first-order Markov process. However, since K is an orthonormal matrix, k (n) k
^t (l) = 0 [l ≠ n] and k (n) k ^t (n) = 1. According to the stationary first-order Markov process, the output from the information source whose correlation coefficient is ρ is r = (r ₁ , r ₂ , r ₃ , ...,
r _N ), the Nth-order covariance matrix R is expressed as in (Equation 1). However, each output r _n from the information source has an average of 0 and a variance of 1.

[0004]

[Equation 1]

A method of deriving the KLT basis vector is shown below. The first row vector k (1) is k (1) Rk
^It is chosen to maximize ^t (1) and k (1) k ^t (1) = 1. The second row vector k (2) is k
(2) Rk ^t (2) maximize, and k (2) k
^It is chosen such that ^t (1) = 0 and k (2) k ^t (2) = 1. The third row vector k (3) is k (3) Rk ^t
Maximize (3), and k (3) k ^t (1) = 0, k
(3) k ^t (2) = 0, k (3) k ^t (3) = 1. Thereafter, similarly, the n-th row vector k (n)
Is selected, and finally the Nth row vector k (N) is k
(N) k ^t (l) = 0 [1 ≦ l <N] and k (N) k ^t
It is chosen such that (N) = 1. It is known that the base row vector k (n) of the KLT thus selected becomes the eigenvector of the covariance matrix R, and k (n) Rk ^t (n) gives its eigenvalue. This means that when performing orthogonal transformation on a signal sequence having a correlation coefficient ρ according to a stationary first-order Markov process, the KLT has the best performance for concentrating energy on a few coefficients. Therefore, KLT is an orthogonal transform that gives the best coding efficiency. It is also well known that when the correlation coefficient ρ of R is set to a value extremely close to 1.0, the basis vector of KLT determined for this R becomes the basis vector of DCT. Therefore, the DCT follows the stationary first-order Markov process, and the KLT is used for coding a signal sequence having extremely high correlation.
Gives the same coding efficiency as.

The conventional image signal encoding method and decoding method use DCT as orthogonal transformation under the assumption that the image signal is stationary and has a very high correlation.
A conventional image signal encoding method using DCT will be described below with reference to the drawings. As shown in FIG. 3, the conventional image signal coding method is composed of a pixel block cutout unit 101, a DCT unit 102, a quantization unit 103, and a variable length coding unit 104.

Each component will be described below. First,
The pixel block cutout unit 101 sequentially extracts N-th order square pixel blocks P from the original image signal. When the inter-frame prediction method is used, the difference value between the previous frame and the current frame is the original image signal, from which the Nth-order square pixel block P is extracted. Next, the DCT means 102 performs an operation of TPT ^{t on the} Nth-order square pixel block P using the Nth-order DCT basis matrix T to calculate an Nth-order square coefficient block C. When the image signal in the pixel block is stationary and has a very high correlation, the DCT concentrates the energy of the pixel block on a small number of coefficients. Further, the quantizing means 103 quantizes each component of the Nth-order square coefficient block C to generate an Nth-order square quantized coefficient block Q. Here, the coefficient having a small value is quantized to zero. Finally, each coefficient of the Nth-order square quantized coefficient block Q is variable length coding means 10
4 is converted into a variable length code. At this time, efficient coding is realized by using only the value of the significant coefficient that is not 0 as the variable length code.

Next, a conventional image signal decoding method using DCT will be described with reference to the drawings. As shown in FIG. 4, in the conventional image signal decoding method, the variable length code decoding means 201, the dequantization means 202, and the inverse DCT means 20 are used.
3 and the pixel block arranging means 204.

Each component will be described below. First,
The variable length code decoding means 201 decodes the variable length code in block units, and generates an Nth-order square quantized coefficient block Q '. Next, the inverse quantizing means 202 inversely quantizes the Nth-order square quantized coefficient block Q ′ to generate an Nth-order square coefficient block C ′. Further, the inverse DCT means 2
03 performs an operation T ^t C'T on the Nth-order square coefficient block C'using the Nth-order DCT basis matrix T to calculate an Nth-order square pixel block P '. Finally, the pixel block arranging means 204 sequentially arranges the Nth-order square pixel blocks P ′ at predetermined positions in the image to obtain a decoded image signal. When the inter-frame prediction method is used, the decoded image signal represents the difference value between the previous frame and the current frame.

Therefore, when the image signals are stationary and have a very high correlation, efficient coding and decoding can be realized by using the DCT as in the conventional example.

[0011]

However, in the above-mentioned conventional image signal encoding and decoding methods using DCT, the image signal in the Nth-order square pixel block for DCT is not stationary but local. When it has a feature, there is a problem that sufficient encoding efficiency cannot be obtained. Furthermore, DC
Since T disperses the energy of a portion having a local feature in the Nth-order square pixel block over a large number of coefficients, quantization of these coefficients causes peculiar noise called mosquito noise to occur in the decoded image signal. There was a problem that caused it.

The present invention solves the above-mentioned conventional problems and is efficient even when the image signal in the N-th square pixel block has a local feature that is not stationary, and suppresses the generation of mosquito noise. It is an object of the present invention to provide an encoding method and a decoding method, and a method for calculating an orthonormal basis matrix used for orthogonal transformation of the encoding and decoding.

[0013]

In order to achieve the above object, a method for calculating an orthonormal basis matrix according to the present invention is a method of calculating N orthogonal basis row vectors b of N orthogonal square orthonormal basis matrices B which are orthogonal to each other. In obtaining (1), b (2), ..., B (N), the nth base row vector b (n) is b (n) Rb ^t (n) with respect to the Nth-order square covariance matrix R. Is maximized and is orthogonal to all b (1), b (2), ..., B (n-1), and in the method of sequentially obtaining from b (1) to b (N), different N The second-order square covariance matrix R (n) is used.

The image signal encoding method of the present invention is
Using the Nth-order square orthonormal basis matrix B calculated by the above method, for the Nth-order square pixel block P, BPB ^t
Then, the orthogonal transformation is performed to obtain the Nth-order square coefficient block C.

Furthermore, the image signal decoding method of the present invention uses the Nth-order square orthonormal basis matrix B calculated by the above method, and B is used for the Nth-order square coefficient block C '.
An orthogonal transformation of ^t C'B is performed to obtain an Nth-order square pixel block P '.

[0016]

In the method of calculating the orthonormal basis matrix of the present invention, when obtaining the base row vector b (n) of the Nth-order square orthonormal basis matrix B, not only the covariance matrix corresponding to the stationary signal sequence but also the non-covariance matrix Since different covariance matrices R (n) such as covariance matrices corresponding to signal sequences having stationary and local characteristics are used, orthogonal transform for concentrating energy to a small number of coefficients for signal sequences of various properties Can generate the basis matrix of.

Furthermore, an image signal encoding method and an image signal decoding method of the present invention are the same as the above Nth-order square orthonormal basis matrix B.
Since the orthogonal transformation is performed by using, even if the image signal in the N-th square pixel block has a local feature that is not stationary, encoding efficiency is good and generation of mosquito noise is suppressed.

[0018]

EXAMPLE A method of calculating an orthonormal basis matrix will be described below as a first example of the present invention. First, as shown in (Equation 2), the Nth-order square orthonormal basis matrix to be calculated is B, and the Nth-order basis row vector is b (n) [1 ≦ n ≦ N]. In addition, the Nth-order square covariance matrix corresponding to b (n) is R
(N).

[0019]

[Equation 2]

First, the first row vector b (1) is
The norm of b (1) is 1, that is, b (1) b ^t (1) =
Under the constraint of 1, we choose to maximize b (1) R (1) b ^t (1). This operation means that when orthogonal transformation is performed on the signal sequence having the statistical property represented by the covariance matrix R (1), the energy concentrated on the coefficient corresponding to b (1) is maximized.

The second row vector b (2) is b
The norm of (2) is 1 and b (2) has already been obtained b
Orthogonal to (1), ie b (2) b ^t (2) = 1,
Under the constraint that b (2) b ^t (1) = 0, b (2) R
(2) b ^t is chosen to maximize (2). This operation means that the energy concentrated on the coefficient corresponding to b (2) is maximized when the orthogonal transformation is performed on the signal sequence having the statistical property represented by the covariance matrix R (2).

The third row vector b (3) is b
The norm of (3) is 1 and b (3) has already been obtained b
(1) and b (2) are orthogonal, that is, b (3) b
^t (3) = 1, b (3) b ^t (1) = 0, b (3) b ^t
Under the constraint (2) = 0, b (3) R (3) b
Selected to maximize ^t (3). This operation means that when orthogonal transformation is performed on the signal sequence having the statistical property represented by the covariance matrix R (3), the energy concentrated on the coefficient corresponding to b (3) is maximized.

Thereafter, similarly, the nth row vector b (n) is selected, and finally the Nth row vector b (N) is b.
The norm of (N) is 1 and b (N) has already been calculated.
(1), b (2), ..., b (N-2), b (N-1) are orthogonal, that is, b (N) b ^t (N) = 1 and b
Only (N) b ^t (l) = 0 [1 ≦ l <N].

The method of selecting each b (n) is based on, for example, nonlinear programming.
The problem is b (n) b^t(N) = 1, b (n) b
^t(L) = 0 [1.ltoreq.l <n] as a constraint condition -b (n) R
(N) b ^tIt can be expressed as a problem of minimizing (n). this
Non-linear programming problems such as, for example, extended Lagrange multipliers
The optimal solution can be obtained using the method.

An example of R (n) [1 ≦ n <4] when R (n) is, for example, a fourth-order covariance matrix (Equation 3)
To (Equation 4). Of these, R (1) and R (2)
Is a covariance matrix corresponding to a stationary signal sequence, and R (3) is a covariance matrix corresponding to a non-stationary signal sequence.

[0026]

(Equation 3)

[0027]

[Equation 4]

According to the method of calculating the orthonormal basis matrix as described above, it is possible to generate the basis matrix of the orthogonal transformation for concentrating the energy in a small number of coefficients with respect to the signal sequence of various properties.

The selection method of each b (n) and the value of R (n) described above are examples, and other selection methods and other values may be used. Next, as a second embodiment of the present invention, an image signal coding method will be described with reference to the drawings. As shown in FIG. 1, the image signal encoding method of the present invention is
Pixel block cutout means 301, orthogonal transformation means 30
2, a quantizer 303 and a variable length encoder 304.

Each component will be described below. First,
The pixel block cutout unit 301 sequentially extracts N-th order square pixel blocks P from the original image signal. When the inter-frame prediction method is used, the difference value between the previous frame and the current frame is the original image signal, from which the Nth-order square pixel block P is extracted. Next, orthogonal transformation means 302
Performs an operation of BPB ^{t on the} Nth-order square pixel block P using the Nth-order basis matrix B calculated by the orthonormal basis matrix calculation method to calculate an Nth-order square coefficient block C. Further, the quantizing means 303 quantizes each component of the Nth-order square coefficient block C to obtain N,
The next square quantized coefficient block Q is generated. Here, the coefficient having a small value is quantized to zero. Finally, each coefficient of the Nth-order square quantized coefficient block Q is converted into a variable length code by the variable length coding means 304. On this occasion,
Efficient encoding is realized by using only the value of the significant coefficient that is not 0 as the variable length code.

Next, as a third embodiment of the present invention, an image signal decoding method will be described with reference to the drawings. As shown in FIG. 2, the image signal decoding method of the present invention is
The variable length code decoding means 401, the inverse quantization means 402, the inverse orthogonal transformation means 403, and the pixel block arrangement means 404 are included.

Each component will be described below. First,
The variable-length code decoding means 401 decodes the variable-length code in block units to generate an Nth-order square quantized coefficient block Q '. Then, the inverse quantizing means 402 inversely quantizes the Nth-order square quantized coefficient block Q ′ to generate an Nth-order square coefficient block C ′. Furthermore, the inverse orthogonal transform unit 403 to the N th square coefficient block C ', performs B ^t C'B comprising calculation using the orthonormal basis matrix N following basis matrix B calculated by the calculation method of the , Nth-order square pixel block P ′ is calculated. Finally, the pixel block arranging means 404 causes the Nth-order square pixel block P '.
A decoded image signal can be obtained by sequentially arranging at the predetermined positions of the image. When the inter-frame prediction method is used, the decoded image signal represents the difference value between the previous frame and the current frame.

Therefore, according to the above-described image signal encoding method and image signal decoding method, orthogonal transformation is performed by the basis matrix for concentrating energy in a small number of coefficients for signal sequences of various characteristics. Even if the image signal has local characteristics instead of stationary, it is efficient and suppresses the generation of mosquito noise.

The above-described image signal encoding method and image signal decoding method are merely examples, and other configurations may be adopted other than the orthogonal transform means 302 and the inverse orthogonal transform means 403.

[0035]

As is apparent from the above embodiments, according to the orthogonal basis matrix calculation method of the present invention, when obtaining the basis vector of the orthonormal basis matrix, only the covariance matrix corresponding to a stationary signal sequence is obtained. However, since different covariance matrices are used for each basis vector, such as a covariance matrix corresponding to a non-stationary, local feature signal sequence, the energy is reduced to a small number of coefficients for signal sequences of various properties. It is possible to generate the basis matrix of the orthogonal transformation to be concentrated.

Further, according to the image signal encoding method and image signal decoding method of the present invention, orthogonal transformation is performed using a basis matrix for concentrating energy in a small number of coefficients for signal sequences of various characteristics. Even if the image signal has local characteristics instead of stationary, the coding efficiency is high and the generation of mosquito noise is suppressed, so that a high-quality decoded image signal can be obtained.

[Brief description of drawings]

FIG. 1 is a diagram illustrating an image signal encoding method according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating a method of decoding an image signal according to an embodiment of the present invention.

FIG. 3 is a diagram illustrating a conventional image signal encoding method.

FIG. 4 is a diagram for explaining a conventional image signal decoding method.

[Explanation of symbols]

 301 Pixel Block Extracting Means 302 Orthogonal Transforming Means 303 Quantizing Means 304 Variable Length Coding Means 401 Variable Length Code Decoding Means 402 Inverse Quantizing Means 403 Inverse Orthogonal Transforming Means 404 Pixel Block Arranging Means

Claims (3)

[Claims] 1. N-th order N-order basis row vectors b (1), b (2), of the N-th order square orthonormal basis matrix B which are orthogonal to each other.
, B (N), the n-th base row vector b (n) is b (n) with respect to the Nth-order square covariance matrix R.
Rb ^t (n) is maximized, and b (1), b (2 ),
, B (n−1) are orthogonal to all b,
An orthonormal basis matrix calculation method characterized in that different Nth-order square covariance matrices R (n) are used in the method of sequentially obtaining from (1) to b (N). 2. Using the Nth-order square orthonormal basis matrix B calculated by the method according to claim 1, an Nth-order square pixel block P is subjected to an orthogonal transformation BPB ^t to obtain an Nth-order square coefficient block C. A method for encoding an image signal, which is characterized by being obtained. 3. An N-th order tetragonal coefficient block C ′ is subjected to an orthogonal transformation of B ^t C′B by using an N-th order square orthonormal basis matrix B calculated by the method according to claim 1, A method of decoding an image signal, which comprises obtaining a pixel block P '.