JPH0711832B2 - High-efficiency image coding device - Google Patents

Description

TECHNICAL FIELD The present invention relates to an image coding method for efficiently coding an image signal.

(Prior Art) As a high-efficiency coding method for image signals, an image is divided into small blocks made up of a plurality of samples, orthogonal transform such as discrete cosine transform is applied to each of the small blocks, and the transform coefficient thereof is applied. A transform coding method for coding and transmitting is known. When a signal having a strong correlation within a block, such as an image signal, is subjected to orthogonal transformation, most of it is concentrated in low-order transform coefficients, and the signal power components are largely attenuated as the order becomes higher. Therefore, conventionally, a low-order transform coefficient is subjected to a fine encoding process, and conversely, a high-order transform coefficient is roughly encoded or not encoded at all, thereby improving the information amount compression effect. . However, by reducing the high-order transform coefficient components in this way, the high-frequency components of the reproduced image are eliminated and the resolution is reduced, so that the deterioration of the reproduced image quality becomes conspicuous. If the compression rate is further increased, the number of transform coefficients that are not subjected to the encoding process increases, so that the image quality is unacceptably deteriorated.

On the other hand, a method of applying a vector quantization method, which is known as an encoding method that can realize an encoding characteristic close to the rate-distortion limit in principle, in an orthogonal transform domain has been announced.
This method utilizes the property that the transform coefficient can be approximated by a standard probability distribution in order to obtain good coding characteristics. That is, in this method, the two-dimensional conversion coefficient F (u, v)
Is divided into bands in which u + v = const, and vector quantization is performed for each band under the assumption that the variance values of transform coefficients in each band are equal. However, in this method, in order to increase the compression rate, it is necessary to exclude the band in which u + v is large from the target of the encoding process, and the vector quantization process must be performed in parallel in band units. There is a drawback that there are restrictions in realizing the device.

(Problems to be Solved by the Invention) An object of the present invention is to reduce deterioration of image quality caused by loss of high-frequency components in such conventional transform coding, and to realize a high-efficiency image code that is easy to realize in the apparatus. It is to provide the system.

(Means for Solving the Problems) In order to achieve the above-mentioned object, the present invention performs orthogonal transformation in block units on an image signal or various prediction error signals for the image signal, and converts a part of transform coefficients in the block. In addition to performing vector quantization, the quantization residual component of the vector quantization is obtained, and a scalar is calculated based on the bit allocation table created by considering statistical properties for the quantization residual component. It is to quantize. Note that vector quantization means that in a space having the number of block elements (in this case, the number of pixels in a block),
It means to determine the subspace in which the block exists. Blocks with the same vector index have similar or statistical properties because they are in the same subspace. For example, if the vector quantized residual component in each element of the image block on which vector quantization is performed is calculated for a sufficiently large number of image blocks existing in the same subspace, and its average power is calculated, the value is Converge. The power of the vector quantization residual component of many image blocks existing in the same subspace has a value close to this average power. Further, the average power of the vector quantized residual component for each subspace has a large correlation with the vector index of that subspace.

That is, the high-efficiency image coding system of the present invention comprises an orthogonal transform means (2) for subjecting an image signal or various prediction error signals to the image signal to orthogonal transform in block units, and transform within a block obtained by the orthogonal transform. Using some of the coefficients,
Vector quantization means (3, 4) for performing vector quantization of the block, and a vector quantization residual by calculating the difference between the output signal of the orthogonal transformation means and the output signal of the vector quantization means for each component. Vector quantization residual calculation means (5) for outputting a signal, and scalar quantization means (6) for performing scalar quantization on the calculated vector quantization residual signal based on a bit allocation table given for each block. ,
7) The basic structure is to have and.

(Operation) The present invention collectively vector-quantizes only a part of transform coefficients obtained by orthogonally transforming each block, for example, a low-order component on which most of signal power is concentrated, and further residuals after vector quantization. Basically, a large number of bits are assigned to the coefficient with a large residual component including the high-order component for the signal, and the scalar quantum is based on the bit assignment table that is predetermined to match the statistical properties. Turn into. In this way, in addition to vector quantization, scalar quantization for the residual signal is also performed, so even if there is a high-frequency component with a large signal power, it is possible to reduce the deterioration of image quality without dropping out and to reduce the device scale. Can be reduced. In addition, it is necessary to increase the block size in order to improve encoding efficiency, but if the block size is increased, the arithmetic circuit and memory for that increase in size,
It is possible to solve the conventional problem that the application of the vector quantization method is practically difficult.

(Embodiment) FIG. 1 is a diagram showing a basic embodiment of the present invention, in which 1 is a signal input terminal, 2 is an orthogonal transformation circuit, 3 is a vector quantization circuit, 4 is a codebook storage circuit, and 5 is subtraction. Reference numeral 6 is a scalar quantization circuit, 7 is a bit allocation table storage circuit, and 8 is a signal output terminal.

The image block composed of N × N pixels input from the signal input terminal 1 is subjected to, for example, two-dimensional discrete cosine transform by the orthogonal transform circuit 2. As for the transform coefficient, as shown in FIG. 2, only the low-order transform coefficient is used to quantize into a representative vector (codeword) in the vector quantization circuit 3 with reference to the codebook storage circuit 4. The codebook storage circuit 4 stores a vector having only the low-order transform coefficients shown in FIG. 2 as a component, and the vector having the minimum distortion from the stored vector is the vector quantization. Selected in the circuit 3, a representative vector is created and output to the subtraction circuit 5. The vector quantization circuit 3 outputs a code indicating the representative vector, that is, an index signal for identifying the representative vector, to the bit allocation table storage circuit 7 and the vector quantized signal output terminal 9. In the bit allocation display storage circuit 7, a bit allocation table is created and stored in advance for each index signal that identifies each representative vector.

The method of creating the bit allocation table will be described. The bit allocation table is created by using the statistical property of the vector quantization residual that the power of the vector quantization residual component of many image blocks existing in the same subspace has a value close to this average power. . That is, a sufficient number of image blocks corresponding to a certain index signal are taken. The residual power in each image block is determined, and when the average residual power for a sufficient number of image blocks is R, the error power after scalar quantization in each image block is determined to obtain a sufficient number of images. Assuming that the average error power for a block is D, the bit allocation number (B bits) for quantizing the residual in that element is calculated by calculating, for example, 2B = log2 (R / D). Such calculation is performed for each element of the corresponding image block for each vector index, and a bit allocation table for each vector index is created.
Then, the subtraction circuit 5 obtains the difference between the input and output vectors of the vector quantization circuit 3. This difference, that is, the residual component is quantized by the scalar quantization circuit 6. The bit allocation to each residual component in the scalar quantization circuit 6 is performed by selecting the bit selected by the vector index output from the vector quantization circuit 3 from the bit allocation table stored in the bit allocation table storage circuit 7. It is based on the allocation table. The output of the scalar quantization circuit 6 is output from the scalar quantization signal output terminal 8.

FIG. 3 is a diagram for explaining the operation principle of the above vector quantization circuit and the scalar quantization circuit of the residual component. First, only the transform coefficient of the low-order component, which is a part of the transform coefficient of the input block (the upper left triangular portion shown in FIG. 3A), is the target of vector quantization and vector quantization is performed. In the quantization, a component that is not the target of vector quantization is added as “0” to the codebook to form a representative vector (FIG. 3 (b)). Next, each component is subtracted from the transform coefficient vector of the input block to obtain the vector quantized residual vector (FIG. 3 (c)). Quantization is performed for each residual vector based on the bit allocation table (FIG. 3 (d)) that is determined so as to match the statistical properties of the residual component.

FIG. 4 is a diagram showing an embodiment in which the present invention is applied to motion compensation interframe coding of a television signal.
1 is a signal input terminal, 102 is a scan conversion circuit, 103 is a motion vector detection circuit, 104 is a variable delay circuit, 105 is a frame memory, 106 is a delay circuit, 107 is a subtraction circuit, 108 is a significant block detection circuit, and 109 is orthogonal. Transform circuit, 110 is a vector quantization circuit, 111 is a subtraction circuit, 112 is a vector quantization residual accumulation circuit, 113 is a code assignment control circuit, 114 is a scalar quantization circuit, 115 and 116 are addition circuits, and 117 is an inverse orthogonal transform. Circuit, 118 delay circuit, 119 multiplexing circuit, 120 buffer memory, 12
Reference numeral 1 is a codebook storage circuit, and 122 is a signal output terminal.

First, the digitized image signal input from the signal input terminal 101 in accordance with the raster scanning order is the scan conversion circuit 102.
According to the predetermined block size such as 4 × 4, 8 × 8, 16 × 16, etc., the blocks are rearranged in the block scanning order. The blocked image signal is input to the motion vector detection circuit 103, and the motion vector indicating the displacement of the subject between frames is detected in block units with reference to the reproduced image of the previous frame stored in the frame memory 105. . The variable delay circuit 104 delays the reference image block from the frame memory 105 used for inter-frame prediction based on the detected motion vector and outputs it to the subtraction circuit 107. On the other hand, the image block from the scan conversion circuit 102 is delayed by the delay circuit 106 for a time corresponding to the sum of the calculation delay time required for motion vector detection and the delay time of the variable delay circuit 104, and the corresponding reference image block in the previous frame. And is input to the subtraction circuit 107 in a state where the delay is adjusted. The subtraction circuit 107 performs the subtraction for each pixel in the block, and the significant block detection circuit 108 uses the subtraction result to select the image block to be encoded / transmitted. For example, for an image block that has a small cumulative sum of absolute values in the block as a result of subtraction and is identified as an invalid block, only the identification code indicating the invalid block is transmitted, and the subsequent encoding process is not applied. The orthogonal transformation circuit 109 performs orthogonal transformation on the significant blocks. The vector quantization circuit 110 performs a block matching process on the significant block with the representative vector (codeword) stored in the codebook storage circuit 121, and a code indicating the representative vector with which the distance between them is minimized. The (vector index) is sent to the multiplexing circuit 119, and the selected representative vector is sent to the subtraction circuit 111. The code indicating the representative vector is also sent to the code assignment control circuit 113. The subtraction circuit 111 performs subtraction in coefficient component units between the input and output vectors of the vector quantization circuit 110, and the vector quantization residual signal of the subtraction result is sent to the scalar quantization circuit 114 at the next stage. At the same time, the vector quantized residual signal is transferred to the residual accumulation circuit 11
2, the residual signal power of each component is accumulated to obtain an accumulated value. Then, by dividing the accumulated value by an accumulator circuit at an appropriate cycle, the average residual power for each component is recalculated and corrected at an appropriate cycle. The average residual power obtained for each vector quantization index and each component is sent to the receiving side through the multiplexing circuit 119 and, at the same time, holds the value until it is recalculated. From the buffer memory 120,
When the amount of data in the buffer memory is acquired and the value is large, the average error power D is corrected largely, and when it is small, the average error power D is corrected small.

Then, the vector quantization index is determined for each block. From the values of the corrected average residual power value R and the corrected average error power D for each corresponding component, for example, 2B
= Log2 (R / D) to modify the number of bits B for scalar quantizing the residual of that component. The scalar quantization circuit 114 quantizes the vector quantized residual signal in coefficient units based on the bit allocation table, and sends a code indicating the quantization level to the multiplexing circuit 119. Also, after adding the representative vector selected by the vector quantization circuit 110 in the addition circuit 115 to the quantized output signal, the orthogonal inverse transform circuit 117 transforms the transformed region into the image region, thereby performing motion compensation for the significant block. The inter-frame difference signal is reproduced. Further addition circuit 116
Then, the motion-compensated reference block in the previous frame is added through the delay circuit 118 having a delay time corresponding to the encoding processing time per block, and the input signal is reproduced. The reproduced image block is stored in the frame memory 105 and used as a motion-compensated inter-frame prediction reference image block of the next frame image.

In the multiplexing circuit 119, the motion vector, the significant / insignificant block identification, the vector index, the bit allocation control code, and the coding indicating the quantization level for each transform coefficient are sent to the time-division multiplexed buffer memory 120.
The buffer memory 120 stores a code generated non-uniformly in block units, and sends the code to the transmission path through the signal output terminal 122 at a constant speed. It also monitors the amount of information stored in the buffer memory and sends a control signal indicating that state to the code allocation control circuit 113.

The vector construction method to which the vector quantization circuit according to the present invention is applied is arbitrary and is not necessarily limited to the example shown in FIG. For example, the DC component F (0,0) has the highest signal power, and in order to improve the encoded image quality, the DC component is separated and finely encoded, but in this case also, F (0 , 0) is excluded from the transform coefficients used for vector quantization, and the number of bit allocations in the scalar quantization circuit of the second stage is increased, it can be easily analogized that the same or higher coding efficiency can be obtained.

(Effects of the Invention) As described above, the present invention utilizes the correlations of image signals to obtain effective vector vectors in the orthogonal transform domain.
The scalar quantization method is configured to enable coding that follows the local statistics of the image signal, and reduce image quality deterioration due to missing high frequency components. Further, since it can be easily applied when it is necessary to increase the block size which is the encoding processing unit in order to increase the compression rate, it is applied to the high compression interframe encoding method for video conference images and the like. There is an advantage that you can.

[Brief description of drawings]

FIG. 1 is a diagram showing a basic embodiment of the present invention. Second
The figure is a diagram showing an example of a low-order component which is a target of vector quantization in the orthogonal transform coefficient. FIG. 3 is a diagram for explaining a scalar quantization method for vector quantized residual signals. FIG. 3A is a transform coefficient of an input block, and FIG.
Is a representative vector, FIG. 7C is a residual of transform coefficients, and FIG. 7D is a diagram showing bit allocation. FIG. 4 is a diagram showing an embodiment when the present invention is applied to a motion compensation interframe coding system. 1 ... Signal input terminal, 2 ... Orthogonal transformation circuit, 3 ... Vector quantization circuit, 4 ... Codebook storage circuit, 5 ...
Subtraction circuit, 6 ... Scalar quantization circuit, 7 ... Bit allocation table storage circuit, 8 ... Signal output terminal, 9 ... Vector quantized signal output terminal, 101 ... Signal input terminal, 102 ... Scan conversion circuit , 103 ... motion vector detection circuit, 104 ... variable delay circuit, 105 ... frame memory, 106 ... delay circuit, 107 ... subtraction circuit, 108 ... significant block identification circuit,
109 ... Orthogonal transformation circuit, 110 ... Vector quantization circuit, 11
1 ... Subtraction circuit, 112 ... Vector quantization residual accumulation circuit,
113 ... Code assignment control circuit, 114 ... Scalar quantization circuit, 115 ... Addition circuit, 116 ... Addition circuit, 117 ... Orthogonal inverse transformation circuit, 118 ... Delay circuit, 119 ... Multiplexing circuit, 12
0 …… buffer memory, 121 …… codebook storage circuit,
122 …… Signal output terminal.

Continuation of front page (51) Int.Cl. ⁶ Identification code Office reference number FI Technical display area H04N 7/137 Z

Claims (4)

[Claims] 1. An orthogonal transform means for performing an orthogonal transform using an image signal or various prediction error signals for the image signal as an input block to obtain a transform coefficient for each component of the block, and vector quantizing the block using the transform coefficient. Vector quantization means for obtaining a representative vector, and vector quantization residual calculation means for calculating a difference between the transform coefficient and the representative vector for each component of the block to obtain a vector quantization residual signal, A high-efficiency image encoding device, comprising: a scalar quantization unit configured to perform scalar quantization on the vector quantized residual signal for each component of the block based on a bit allocation table corresponding to the representative vector. . 2. The vector quantizing means stores codebook storing means for storing a codebook having a vector containing only low-order transform coefficients as components, and distortion between the codebook and an input block vector is generated. A vector that selects the smallest vector, adds components of the transform coefficient other than the low-order transform coefficient in the block as "0" and outputs it as a representative vector signal, and identifies the representative vector signal. A high-efficiency image coding apparatus according to claim 1, further comprising: a unit that outputs an index signal. 3. A means for storing the bit allocation table for each representative vector, the scalar quantization means storing a bit allocation table for each representative vector, the bit allocation table being determined so as to match the statistical property of the vector quantized residual signal for each representative vector. Means for selecting a bit allocation table based on the vector index signal, and means for scalar quantizing the vector quantization residual signal by bit allocation for each component of the block according to the bit allocation table. The high-efficiency image coding apparatus according to claim (2), characterized in that. 4. A means for storing the bit allocation table for each representative vector, the scalar quantization means storing a bit allocation table for each representative vector, which is determined so as to match the statistical property of the vector quantized residual signal for each representative vector. Means for selecting a bit allocation table based on the vector index signal, means for modifying the bit allocation table based on the vector quantized residual signal, and according to the bit allocation table for each component of the block A means for scalar-quantizing the vector quantization residual signal by bit allocation, and the high-efficiency image encoding device according to claim (2). (5) The scalar quantization means stores the scalar-quantized signal in a buffer memory and generates a control signal indicating the monitoring state of the stored information amount; and the vector quantized residual signal for each representative vector. Is determined to match the statistical properties of, means for storing a bit allocation table for each representative vector, means for selecting a bit allocation table based on the vector index signal, of the vector quantization residual signal Means for accumulating power and outputting an accumulated value, means for modifying the bit allocation table based on the control signal and the accumulated value, and according to the bit allocation table for each component of the block A means for scalar-quantizing the vector quantization residual signal by bit allocation, and a high-efficiency image coding apparatus according to claim (2). Place