JP2691616B2 - Discrete cosine transform coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Outline] Regarding a discrete cosine transform coding method for high-efficiency coding of an input image signal, a high-frequency component that does not cause visual deterioration of a reproduced image is omitted to further improve the performance. A discrete cosine transform unit that performs a discrete cosine transform on a block of a plurality of pixels of an input image signal for the purpose of efficient coding, and a quantizer that quantizes a transform output signal of the discrete cosine transform unit. ,
A method for performing discrete cosine transform coding of the input image signal, comprising a coding section for coding the quantized output signal of the quantizing section, and detecting a function indicating a high frequency component in the block. A function detecting unit is provided, the quantized output signal of the quantizing unit indicating a high frequency component is forcibly converted into 0 based on the function from the function detecting unit, and the signal not including the 0 is converted into the encoding unit. It is configured to be encoded in.

[Industrial applications]

The present invention relates to a discrete cosine transform coding method for high efficiency coding of an input image signal.

In videoconferencing, videophone, etc., various high-efficiency coding schemes for performing band compression of image signals have been proposed.
Recently, discrete cosine transform (Discrete C
osine transform) is attracting attention. This discrete cosine transform is one in which a plurality of pixels are set as one block and is orthogonally transformed into transform coefficients by a transform matrix. Generally, the transform coefficients are concentrated on the low frequency component side, although it varies depending on the property of the image. Become. Then, this transform coefficient is entropy-coded and transmitted, and on the receiving side,
The reproduced image signal is obtained by performing entropy decoding, inverse quantization, and inverse discrete cosine transform.

In such a discrete cosine transform coding method, there is a demand for further improving the efficiency.

[Conventional technology]

FIG. 10 is a block diagram of a conventional example, and the transmission unit includes a discrete cosine transform unit 30, a quantization unit 31, a scanning unit 32, and an encoding unit 33.
The receiving unit includes a decoding unit 34, an inverse scanning unit 35, an inverse quantization unit 36, and an inverse discrete cosine transform unit 37.

The input image signal is a signal obtained by digitizing an image signal from a television camera or the like or a signal obtained by performing an inter-frame difference process on the image signal, and one of the input image signals.
For example, as shown in FIG. 11, the screen portion is divided into blocks of N × N pixels, and the pixels f (u, v) of each block are divided.
Then, the discrete cosine transform unit 30 performs discrete cosine transform, the transform coefficients F (i, j) are individually quantized by the quantizing unit 31, and the scanning unit 32 scans to quantize the two-dimensional array. The encoded output signal is made into a one-dimensional array and sequentially entropy-encoded in the encoding unit 33 and transmitted to the receiving side.

On the receiving side, entropy decoding is performed in the decoding unit 34, processing reverse to that of the scanning unit 32 on the transmitting side is performed in the reverse scanning unit 35, and the quantization unit 31 is processed in the inverse quantization unit 36. The inverse discrete cosine transform unit 37 performs inverse quantization of the process, and the inverse cosine transform unit 37 performs the inverse process of the discrete cosine transform unit 30 to reproduce an image signal, which is added to a display device (not shown).

The discrete cosine transform in the discrete cosine transform unit 30 is
It is expressed by the following equations (1) and (2).

Using the equations (1) and (2), the pixel f
When (u, v) is converted to F (i, j), the conversion coefficient F (i, j)
j) indicates a frequency component corresponding to the value of i, j. The smaller the value of i, j, the lower the frequency component, and F (0,0)
Indicates a direct current component. For example, N on the screen in FIG.
In a block consisting of × N pixels, if N = 8,
The shaded block is shown in an enlarged form below, and by performing the discrete cosine transform, the arrow a direction becomes a low frequency component and the arrow b direction becomes a high frequency component. In the case of a general screen, it is known that the low-frequency components are concentrated by performing the discrete cosine transform. That is, many of the high frequency components in the direction of the arrow b become 0.

The conversion coefficient F (i, j) corresponding to the block is quantized in the quantization unit 31 and scanned in the scanning unit 32.
In the lower block of FIG. 11, c is a zigzag scan, d is a horizontal scan, and e (dotted line) is a vertical scan. Toward or to the opposite direction, the scanning unit 32 scans, and in addition to the sequential encoding unit 33, entropy encoding is performed.

As described above, since the high-frequency component is often 0, scanning is performed from the high-frequency region side, and an end-of-bound code EOB (End Of Bound) is placed at a position where it is no longer 0.
Is added. For example, for the quantized signal of the block of 8 × 8 pixels shown in FIG. 12, zigzag scanning is performed from the high frequency region side, and the position (value of −7) that is no longer 0
Add the end-of-bound code EOB. Thereby,
When encoding and transmitting from the low frequency region side, regarding the high frequency component after the position of the end of bound code EOB,
Since all of them are 0, it is not necessary to transmit them, and the coding efficiency can be improved.

On the receiving side, the end of the coded signal of the block can be identified by the end-of-bound code EOB, and the decoding process can be performed.

[Problems to be solved by the invention]

In the above-mentioned discrete cosine transform coding method of the conventional example, for example, in the case of the block shown in FIG. 12, when the zigzag scanning is performed from the low frequency region side to code and transmit, 103
1, -9, -5,3,20,15,0, ... 0, -7 (EOB), so even if 0 after the end of bound EOB can be omitted, end of bound EOB Since it contains many 0s before, it was not possible to sufficiently improve the coding efficiency.

It is an object of the present invention to achieve higher efficiency coding by omitting high frequency components that do not cause visual deterioration of a reproduced image.

[Means for solving the problem]

The discrete cosine transform encoding method of the present invention forcibly sets the high-frequency component, which does not cause a visual problem, to 0 based on the function indicating the high-frequency component of the input image signal. Referring to FIG. explain.

A discrete cosine transform unit 1 that performs a discrete cosine transform on a block of a plurality of pixels of an input image signal, a quantizer 2 that quantizes a transformed output signal of the discrete cosine transform unit 1, and a quantizer 2 An encoding unit 3 for encoding the quantized output signal and a function detecting unit 4 for detecting a function indicating the amount of a high frequency component corresponding to the variation between pixels in the block of the input image signal are provided. When the function from the detection unit 4 indicates that there are few high frequency components in the block of the input image signal, the quantized output signal of the quantized output signal of the quantizer 2 is forced to 0. The quantized output signal is converted and the quantized output signal not including 0 is encoded in the encoding unit 3.

[Action]

Similar to the conventional example, the discrete cosine transform unit 1 performs a discrete cosine transform on an input image signal corresponding to a block made up of a plurality of pixels, quantizes it by a quantization unit 2, and encodes it by an encoding unit 3.
Then, it is entropy coded and transmitted. Further, the function detection unit 4 detects a function indicating a high frequency component due to an edge or a standard deviation in a block of the input image signal, and based on this function, the quantization unit 2 or the encoding unit 3 indicates a high frequency component. The converted output signal is forcibly converted to 0. For example, when a high frequency component of the quantized output signal appears even though the number of edges in the block is small, there is no visual problem even if the high frequency component does not exist. Therefore, such high frequency components that do not cause visual problems are forced to 0.
And reduce the amount of information to be encoded by the encoding unit 3,
It is possible to improve the coding efficiency. Also, the receiving side can reproduce the image signal with the same configuration as the conventional example.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 2 is a block diagram of an embodiment of the present invention, in which the transmitting side has a discrete cosine transform section 10, a quantizing section 11, a zero transforming section 12, a scanning section 13, and an encoding section 14. , And a function detection unit 15. In addition, the receiving unit includes a combining unit 16 and a reverse scanning unit 17
And an inverse quantization unit 18 and an inverse discrete cosine transform unit 19.

With N × N pixels of the input image signal as one block, the discrete cosine transform unit 10 performs discrete cosine transform for each block, and the transform coefficient F of the output of the discrete cosine transform unit 10
(I, j) is quantized by the quantization unit 11 and added to the zero conversion unit 12. Further, the function detecting unit 15 detects a function indicating the amount of the high frequency component corresponding to the variation between pixels in one block of N × N pixels of the input image signal. For example, when edge detection is performed, the amount of change between pixels at the edge portion is large, and the larger the number of edges, the more high-frequency components.

In the zero conversion unit 12, when the function from the function detection unit 15 indicates that the high frequency component in the block of the input image signal is small, the quantization output on the high frequency region side in the quantized output signal of the quantization unit 11 Force the signal to convert to 0. The scanning unit 13
The quantized output signals in the two-dimensional array state are one-dimensionally arrayed by, for example, zigsag scanning and input to the encoding unit 14. The encoding unit 14 entropy-encodes the scan output signal of the scanning unit 13. At this time, 0 is not encoded including the signal of the quantized output signal which is forcibly converted to 0.

Further, as in the conventional example, the receiving unit entropy-decodes the received coded signal by the decoding unit 16, converts the one-dimensional array by the inverse scanning unit 17 into a two-dimensional array, dequantizes it by the inverse quantization unit 18, and inverses it. The discrete cosine transform unit 19 performs an inverse process of the discrete cosine transform to obtain an output image signal.

FIG. 3 and FIG. 4 are flowcharts showing the operation of the function detecting unit 15, showing the case of edge detection by Laplacian. Laplacian mask M (3,3) (center 4,
(Upper, lower, left, and right are set to -1, and the sum of mask coefficients is 0)
, A block f (8,8) consisting of 8 × 8 pixels, a block f ′ (8,8) after processing the mask M, and a threshold TH
And the number of edges COUNT are determined, and the Laplacian mask M is applied as f (I, J) excluding the peripheral pixels of the block. That is, with I = 2, J = 2, the Laplacian mask M is applied to the pixel of f (2,2). By applying this Laplacian mask M, for example, when all the pixels in the block have the same value and the high frequency component is 0,
If all the pixels in the block are 0, and if there is a difference between pixels such as a boundary line in the block, the pixel corresponding to the boundary line does not become 0 because a high frequency component is included.

Then, J = J + 1 is set, it is determined whether J> 7,
If J> 7 is not satisfied, the Laplacian mask M is applied to the pixel of f (I, J + 1). When J> 7, I = I + 1, it is determined whether I> 7, and I> 7.
If not, move to step. Therefore, 8 × 8
Laplacian mask M for 6 × 6 pixels
Will be applied.

Next, set the number of edges COUNT = 0 (see FIG. 4), I
= 2, J = 2, it is determined whether or not the absolute value ABS of the block f ′ (I, J) on which the Laplacian mask M is applied is equal to or more than the threshold value TH. For example, if the pixel can take values from 0 to 255, the threshold TH can be 50. If the value of the block f ′ (2,2) is greater than or equal to the threshold value TH, the number of edges COU
Increase NT by 1 and set J = J + 1. If it is not equal to or more than the threshold value TH, J = J + 1. Then, it is determined whether or not J> 7. If J> 7 is not satisfied, the process proceeds to step. If J> 7, I = I + 1 is set, and it is determined whether or not I> 7. If it is not 7, the process proceeds to step, and if I> 7, the process ends.

By the above-described processing, the number of pixels equal to or larger than the threshold value TH in one block on which the Laplacian mask M is applied is counted as the edge number COUNT. The fact that the number of edges COUNT is large indicates that there are many high frequency components.

FIG. 5 is a flow chart showing the operation of the zero conversion unit 12, in which the parameter U is obtained from the number of edges COUNT and the high frequency component in one block is forcibly set to 0 based on the parameter U. A block of 8 × 8 pixels quantized output signal F ′ (8,8), a matrix S (8 × 8) obtained by expanding this in one dimension, and a parameter U = indicating how many high-frequency components are zero MAX [(6 × 6-COUNT + α),
0] and a matrix MAP (8,8) indicating the scanning order.

The parameter U is for selecting the larger one of (6 × 6-COUNT + α) and 0, and α is a value that can be set arbitrarily, and can be set to 10, for example. Further, the number of edges COUNT is not more than 6 because the number of pixels on which the Laplacian mask M is applied is 6 × 6.

The expansion of one block of the quantized output signal F '(8,8) into a one-dimensional matrix S (K) can be performed, for example, according to the flowchart shown in FIG. That is, F '
(8,8), S (8 × 8) and matrix MAP (8,8) are defined (a), K = 1 of S (K) is set (b), I = 1 is set (c), J = 1 is set (d) as the first position of the matrix, and it is determined whether the scanning order of the matrix MAP (I, J) and the number of the matrix S (K) are the same (e).

When MAP (I, J) = K, the matrix S (K) has F ′ (I,
The value of J) is inserted (f) and J = J + 1 (g). Therefore, the value of F '(1,1) corresponding to the position is inserted into MAP (1,1) of scanning order 1 in S (1), and the scanning order 2 is inserted in S (2).
The value of F '(1,2) corresponding to the position of MAP (1,2) of the above is inserted, and the value of MSP (1,3) of the scanning order 6 is inserted into S (6).

When MAP (I, J) ≠ K, J = J + 1 (g). Next, it is judged whether or not J> 8 (h). If J> 8 is not satisfied, the process proceeds to step (e). If J> 8, I = I + 1.
(I), and it is determined whether I> 8 (j). When I> 8 is not satisfied, the process proceeds to step (d), J = 1 is set, and then the above-described operation is repeated. Further, when I> 8, K = K + 1 is set (k), and it is determined whether K> 64 (1). When K> 64 is not satisfied, the process proceeds to step (c), and after I = 1, the above-mentioned operation is performed. repeat. When K> 64, it means the end of the matrix S (K), so the process is terminated.

By the above-described processing, the one-block two-dimensional array of the quantized signal F ′ (I, J) is converted into the one-dimensional array matrix S (K) according to the zigzag scanning. Since the high frequency components are distributed on the tail side of the one-dimensional array, the process of forcibly setting the value to 0 according to the parameter U is performed from the tail of the one-dimensional array. That is, in FIG. 5, the counter C is set to 0, K = 64, C = C + 1, and compared with the parameter U. When U> C, S (K) = 0. That is, it is forcibly set to 0. Then, K = K−1 is set, and it is determined whether or not K <1. If K <1 is not satisfied, the process proceeds to step, and if K <1, the process is ended. Also U <C
If it is not, the value of 0 of the high frequency component is larger than the value of the parameter U including the value that is forcibly set to 0, and thus the process of forcibly setting the high frequency component to 0 ends.

Therefore, in the case shown in FIG. 12, (-7) with EOB added can be forcibly set to 0, so that F '
It becomes 0 before the position of (3,3), and the coding efficiency can be improved.

In the above-mentioned embodiment, the 3 × 3 Laplacian mask M is applied to the 8 × 8 pixel block shown by the solid line in FIG. 7 except for the pixels at the block boundaries. Therefore, the Laplacian mask processing is performed on the 6 × 6 pixels indicated by the dotted line. Even when the edge detection is not performed for the pixel at the boundary of such a block, the parameter U indicating the high frequency component as described above is obtained, and the high frequency component is forcibly set to 0, without deteriorating the reproduced image quality. The coding efficiency can be improved.

When the Laplacian mask M including the boundary pixels is applied to the 8 × 8 pixel block indicated by the solid line,
It suffices to apply the Laplacian mask M to the 10 × 10 pixels shown by the chain line, and the Laplacian mask M can be applied to all the pixels.

FIG. 8 is a block diagram of a main part of another embodiment of the present invention, in which 20 is a discrete cosine transform unit, 21 is a quantizer unit, 22 is a zero transform unit, 23 is a scanning line, 24 is an encoding unit, and 25. Is a standard deviation detector. This embodiment corresponds to the case where the function detecting section 15 in the above-mentioned embodiment is the standard deviation detecting section 25, and the property of including a high frequency component is utilized as the standard deviation σ is larger.

FIG. 9 is a flow chart showing the operation of the standard deviation detection unit 25. One block of the input image signal is f (I, J), the standard deviation is σ, and the average value is H (1). First, H =
0, I = 1, J = 1 (2) to (4), average value H = H + f
Find (I, J) (5). Next, set J = J + 1 (6),
It is determined whether or not J> 8 (7). If J> 8 is not satisfied, the process proceeds to step (5). If J> 8, I = I +
It is set to 1 (8), it is determined whether I> 8 (9), and I>
If it is not 8, the process proceeds to step (4), and if I> 8, the process for 8 × 8 pixels is completed, so H = H / 64
Is calculated (10) to obtain the average value H.

Next, the standard deviation σ is set to 0 (11), I = 1 and J = 1 (1
2), (13), σ = σ + [f (I, J) −H] ² is calculated (14), J = I + 1 is set (15), and it is determined whether J> 8 (16) , If J> 8, go to step (14). If J> 8, set I = I + 1 (17). Determine if I> 8 (18). If I> 8, go to step. The process moves to (13), and if I> 8, the processing for 8 × 8 pixels is completed. Is calculated (19) to obtain the standard deviation σ.

Based on this standard deviation σ, the zero conversion section 22 forcibly sets the high frequency component in the quantized output signal to 0, and as the parameter U in the above-mentioned embodiment shown in FIG.
U = MAX [β−σ + α, 0]. Β in this case is the maximum value that α can take. Step again (19)
It is possible to set β and α correspondingly by using the standard deviation σ of the value obtained by omitting the processing of the square root in.

In the above-described embodiment, the case where one block is 8 × 8 is shown, but it is also possible to have another pixel number configuration. Moreover, even if the high frequency component of 0 is encoded and transmitted,
The transmission efficiency can be improved by reducing the number of assigned codes.

〔The invention's effect〕

As described above, according to the present invention, the function detection unit 4 is provided to detect a function based on the number of edges or standard deviation indicating a high frequency component corresponding to a block of an input image signal, and based on the function, the discrete cosine transform is performed. The high frequency component of the quantized output signal obtained by quantizing the output is forcibly set to 0. Therefore, the coding efficiency is improved by finding a high frequency component that does not cause a visual problem and forcibly setting it to 0. There is an advantage that can be done.

[Brief description of the drawings]

FIG. 1 is an explanatory view of the principle of the present invention, FIG. 2 is a block diagram of an embodiment of the present invention, FIGS. 3 and 4 are operation flowcharts of a function detecting unit of an embodiment of the present invention, and FIG. Is an operation flowchart of the zero converter of one embodiment of the present invention, FIG. 6 is a flowchart of a scanning operation, FIG. 7 is an explanatory diagram of a block,
FIG. 8 is a block diagram of essential parts of another embodiment of the present invention, FIG. 9 is a flow chart of another embodiment of the present invention, FIG. 10 is a block diagram of a conventional example, and FIG. 11 is a block scanning explanatory diagram. , 12th
The figure is a scanning explanatory view of a conventional example. Reference numeral 1 is a discrete cosine transform unit, 2 is a quantization unit, 3 is an encoding unit, and 4 is a function detection unit.

Claims (1)

(57) [Claims] 1. A discrete cosine transform section (1) for performing a discrete cosine transform on a block of a plurality of pixels of an input image signal, and a quantizing section for quantizing a transform output signal of the discrete cosine transform section (1). (2) and an encoding unit (3) for encoding the quantized output signal of the quantization unit (2),
In the method of performing the discrete cosine transform coding of the input image signal, a function detecting section (4) for detecting a function indicating the amount of high frequency component corresponding to the variation between pixels in the block of the input image signal. And when the function of the function detector (4) indicates that the high frequency component in the block of the input image signal is small,
The quantized output signal in the high frequency region of the quantized output signal of the quantizing unit (2) is forcibly converted to 0, and the quantized output not including 0 in the encoding unit (3). Discrete cosine transform coding method characterized by coding signals.