JPH058909B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an image signal encoding device, a decoding device, and a method thereof for shortening the transmission time of image signals or reducing the storage capacity.

(Prior Art) Data compression methods for multivalued images (for example, 8 bits per pixel, 256 levels) include information preservation type encoding and information non-preservation type encoding. Information-preserving encoding does not include quantization in the encoding process, and it is possible to reproduce an image that is exactly the same as the original image through the encoding and decoding processes. ,
High compression ratio cannot be obtained. On the other hand, non-information-preserving encoding refers to one that includes some kind of quantization processing during the encoding process, and the reproduced image contains quantization noise due to the encoding/decoding process, resulting in deterioration of image quality. However, a high compression ratio can be obtained. In the case of non-information preserving encoding, it is generally evaluated based on the relationship between quantization distortion (S/N ratio) and data compression rate (amount of information).
One method for realizing a good relationship between the S/N ratio and the amount of information is a method of vector quantizing transform coefficients after orthogonal transform. A typical example of this method is discrete cosine transform vector quantization (DCT-
VQ) is known. (Reference (1): “Discrete Cosine Transform Vector Quantization (DCT-VQ)” Aizawa,
Harashima, Miyagawa, Journal of the Television Society Vol.39, No.10,
(1985, P920-925) This method performs discrete cosine transform (DCT) on the image signal and then performs vector quantization, which makes it possible to use common statistical properties of images as targets for vector quantization. It has the feature of being able to design a highly versatile vector quantizer. In the DCT-VQ method, an image is divided into blocks of n pixels vertically and n pixels horizontally (n×n pixels), DCT transform is performed in units of blocks, and the obtained n×n transform coefficients are vector quantized. In this case, for example, if n = 8, the number of transform coefficients in one block is 64, and performing 64-dimensional vector quantization using this as one vector increases the encoding processing time. This leads to an increase in the memory capacity of the codebook that is inserted, making it difficult to implement. Therefore, in the DCT-VQ method, the 8×8 transform coefficients are
As shown in the figure, it is divided into multiple bands and a single vector is created using the transform coefficients that fit into one band. (Number of vectors is 14) With this method, the dimension of the vector (the number of components in the vector) is at most 8, which is easy to implement.

(Problems to be Solved by the Invention) The method of dividing the bands in FIG. 2 was determined by considering the following two points.

(1) When the variances of components in a vector are significantly different, uniform information allocation by vector quantization is inefficient.

(2) Generally, the magnitude of the variance of DCT transform coefficients is approximately isotropic from low-order coefficients to high-order coefficients.

However, this property of the second surface holds true when there is no edge component with directionality (vertical, horizontal, diagonal, etc.) in the flat part of the image or in the image within the block. When an image within a block includes edges in the vertical, horizontal, and diagonal directions, as shown in FIG. 3, the distribution of transformation coefficients having large values differs as follows. (The shaded area in the image area in Figure 3 indicates the edge direction, and the shaded area in the transformation area indicates the area that includes a large value of transformation coefficient.) Γ If there is an edge component in the vertical direction, it is A large conversion factor occurs (Figure 3a).

Γ If there is an edge component in the horizontal direction, a large conversion coefficient occurs in the vertical direction from the low order (Fig. 3b).

Γ If there is an edge component in the diagonal direction, a large conversion coefficient occurs from the low order to the diagonal direction (the third
Figure c).

For this reason, when the band division method shown in Figure 2 is applied to the DCT transform coefficients for an image block containing directional edge components, the variance values of the components within the vector will differ greatly, and uniform information allocation by vector quantization will not be possible. Efficiency decreases. Furthermore, as a result, when achieving a high compression rate, the edges of the decoded image become blurred. The present invention solves this problem in the prior art by adaptively switching the band division method so that the variance values of the components in the vector are as equal as possible according to the distribution of the value of the DCT transform coefficient. It is something.

Note that the relationship between the image area and the transformation area shown in FIG. 3 is described in detail in reference document (2).
(Reference (2) “Orthogonal transform coding method with zone adaptation” Miyahara, Tanaka, Hotta, Nakagawa, Journal of the Television Society Vol. 39, No. 10, 1985, pp 898-904) (Example) The following drawings Examples of the present invention will be described with reference to . FIG. 1a is a block diagram showing an example of an encoding device that implements the image signal encoding/decoding method of the present invention, and FIG. 1b is a block diagram showing an example of a decoding device.

In the following explanation, a two-dimensional discrete cosine transform is used as the orthogonal transform, but it is also possible to use other orthogonal transforms such as Hadamard transform.

First, in the encoding device, an image signal is read out in units of blocks by a block readout section 1, which performs DCT transformation. For example, an image signal of 8 bits per pixel is read out with a total of 64 pixels (8 pixels vertically and 8 pixels horizontally) as one block. The read image signal 10 for one block is subjected to a two-dimensional discrete cosine transform in the DCT transform unit 2, and is converted into 8×8
DCT transform coefficients 11 are calculated.

This DCT conversion coefficient 11 is classified into four categories depending on where a large coefficient value exists in the category analysis unit 3, and the corresponding category number 12
is output to. Here, category 1 corresponds to Figure 3 a, where large coefficients are distributed in the horizontal direction from low orders, and category 2 corresponds to Figure 3 b, where large coefficients are distributed in the vertical direction from low orders, Category 3 corresponds to FIG. 3c, where large coefficients are distributed diagonally from the low order, and category 4 corresponds to blocks with no particular directionality, where large coefficients are distributed from the low order. Next, category analysis section 3
We will explain how to find the category number in . FIGS. 4a to 4d show four examples of mask patterns stored in the mask pattern storage section 4.

Category analysis section 3 selects mask patterns from 1 to
Read up to 4 one by one, and calculate the sum of the absolute values of the conversion coefficients at the positions corresponding to 1 in the mask pattern to 1.
The value Ki (average absolute value of the conversion coefficients at the position corresponding to 1) divided by the number of mask patterns i (i = 1 to
4) and outputs the mask pattern number i corresponding to the maximum value as category number 12. Since the position of 1 in each mask pattern corresponds to the position where large coefficient values in each category are distributed, using this analysis method enables category analysis based on the distribution of the values of the conversion coefficients.

The reason why the DC component is set to 0 in each mask pattern in FIG. 4 is because the DC component is unrelated to the directionality of edges and should be excluded from the calculation of Ki used in category analysis.

Another method of category analysis is to set a threshold value T in advance for the conversion coefficients, and calculate the number of conversion coefficients that have a mask pattern of 1 and an absolute value larger than the threshold value T for each category. You can also choose a category that has a large number of items.

In addition, a representative one for each category in advance.
Conversion coefficient Fk (i, j) for block (k is category number, (i, j) indicates position within block)
is determined, this representative conversion coefficient is stored in the mask pattern storage unit 4 instead of the mask pattern, the distance between the input conversion coefficient G(i, j) and the representative conversion coefficient of each category is determined, and this distance is You can also select the smallest category.
An example of how to obtain the distance L _k for the category k is shown in the following equation.

L _k = 〓 i 〓 j√{(,)−(,)} ^2To find the representative transformation coefficients, first calculate the DCT transformation coefficients for each block using multiple test images, and calculate the number of transformation coefficients in the block. Create a training sequence as a vector (64 dimensions for an 8x8 block), use the vector quantization codebook creation method to find representative vectors for the number of categories, and convert them into representative transformation coefficients. It can be done.

Next, in the band dividing section 5, the DCT transform coefficient 11
is divided into a plurality of bands according to category number 12, and a vector 13 whose components are exchange coefficients included in each band is generated. Band table storage section 6
A band table is stored in which the conversion coefficients corresponding to each category are included in one band, each having the same variance value as possible. Examples of band tables corresponding to categories 1 to 4 are shown in FIGS. 5a to 5d. Each number in the band table represents the number of the band to which the conversion coefficient at the corresponding position belongs. Band table 1 has bands divided horizontally, band table 2 has bands divided vertically, band table 3 has bands divided from upper left to lower right, and band table 4 has bands divided from upper right to lower left. These correspond to the distribution of conversion coefficient values belonging to each category.
Band Φ is a band containing only DC components in DCT transformation.

The band dividing unit 5 reads the band table with the corresponding number according to the category number 12 from the band table storage unit 6, and selects the DCT transformation coefficients 11 corresponding to the positions having the same band number in the band table in the order of the band number. , a vector 13 having these as components is generated and sent to the vector quantizer 7. For example, from FIG. 5, the dimension of the vector corresponding to category 1 band number 2 is 8.

The code book storage section 8 stores code books optimally designed in advance for each band within each category. A codebook is a collection of representative vectors (finally output vectors) in vector quantization. For example, for category 1, nine code books are prepared in advance, corresponding to bands 0 to 8, as shown in FIG. Vector quantizer 7
Then, the category number 12 and the hand number of the vector inputted from the band dividing unit 5 (as mentioned earlier, the vectors are sent from the band dividing unit 5 in the order of the band numbers, so the vector quantizer 7 calculates the hand number of the input vector. (You can know the band number.)
The corresponding codebook is read out from the codebook storage unit 8, the one with the shortest distance to the input vector is selected from among the plurality of representative vectors, and the number assigned to that representative vector is output as the vector index 14. do.

To create a codebook, vectors selected from many test images or vectors generated assuming the distribution of each coefficient (for example, Laplace distribution) are used as a training sequence for the LBG algorithm (Reference (3): "Unalgorithm format"). Vector Quantizer Design (An
Algorithm for Uector Quantizer Design)
Y Linde, A. Buzo, RMGray, IEEE
Trans., COM-28, 1, pp.84-95, Jan.
(1980). In addition, the number of representative vectors in each codebook M = 2 ^m is determined in advance by an appropriate value from the variance value of the component included in each band in each category obtained from the training sequence and the target compression ratio. be able to. (Reference (1)) The category number 12 obtained by the category analysis section 5 and the vector index 14 obtained by the vector quantizer 7 are encoded by the encoding section 9 and are coded 1.
5 is output. When an equal length code is used as the code, if the number of categories is 4, the category number can be encoded with 2 bits for each block. In addition, the number of bits allocated to the vector index differs for each category and each band, and the number of representative vectors of the corresponding codebook is set as M (log ₂ M).
Encoded in bits.

Furthermore, if the frequency of occurrence of each category number or each vector index varies, the amount of code can be reduced by using an unequal length code such as a Huffman code.

Next, the decoding device will be explained according to the block diagram of FIG. 1b. The code 15 received from the encoding device is decoded by a decoding section 20, and a category number 31 and a vector index 32 are output.
These are the same as category number 12 and vector index 14 in FIG. 1a, respectively. Next, the vector dequantizer 21 uses the decoded category number 31 and the band number of the vector to be dequantized. can be generated in the vector inverse quantizer 21)
A corresponding codebook is read out from the codebook storage section, and a representative vector that is decoded from among the representative vectors in this codebook and indicated by the vector index is output as a decoded vector 33. Codebook storage unit 22 of the decoding device
The codebook in the codebook is the same as the codebook in the codebook storage section 8 in the encoding device.

The block synthesis unit 23 reads out the band table of the corresponding category from the band table storage unit 24 according to the category number 31, and arranges the input decoded vectors in the order of the band numbers at the corresponding positions in the block indicated by the band table. To go. When decoding and arrangement of all vectors in one block are completed, decoded transform coefficients 34 for one block are output from the block synthesis section 23.

The hand table stored in the band table storage section 24 of the decoding device is the same as the band table stored in the band table storage section 6 of the encoding device. The inverse DCT transform section performs inverse DCT transform on the decoded transform coefficients in units of one block, and sequentially outputs one block worth of decoded images 35. The decoded image signal 35 in units of blocks is finally output through the block writing section 26 to an image output device such as an image storage device or a printer.

In the above explanation, the number of categories was explained as four, but it can be increased or conversely reduced to about two. Block size is also 8x
8, but other sizes may be used. In this example, vector quantization simply finds a representative vector close to the input vector, but it calculates a normalization coefficient from the average variance value of vector components for each band in each category, and uses this normalization coefficient to It is also possible to obtain a decoded vector by dividing each component and then vector quantizing the vector, and after inverse vector quantizing in the decoding device, multiplying each component of the vector by a normalization coefficient. (References (1))
However, in this case, it is necessary to encode the normalization coefficient as additional information for each band of each category.

Furthermore, in this example, the number of representative vectors in the codebook used for each band in each category M = 2 ^m (m indicates the number of bits allocated to the vector)
is set constant in advance, but it can also be adaptively switched depending on the image to be compressed or the target compression rate.

In this method, multiple codebooks are prepared in advance with different numbers of representative vectors for each category and each band, and when the image to be compressed is divided into each category and each band, each vector component in the band is This method adaptively changes which codebook to use based on the average variance value and target compression rate, and this is described in detail in Reference (1) as ``optimal bit allocation.'' When this method is used, it is necessary to encode the number of representative vectors (or the number of bits representing this number) in the codebook used for each category and each band as additional information.

In addition, in order to perform the normalization processing and adaptive bit allocation to the vectors described above more appropriately, the blocks included in each category are further divided into multiple blocks (for example, 4
It is also possible to divide the data into two classes and switch the normalization coefficient and bit allocation for each class. In this case, it is necessary to encode the class number as additional information in addition to the category number for each block. Details of this classification are described in the following reference (4). (Reference (4): “Adaptive coding of monochrome and color images”
and color images), IEEE TRANSACTIONS ON
COMMUNICATIONS) magazine, November 1977 issue,
(pp. 1285-1292) Finally, the DC component (band Φ) in the DCT transform coefficients can also be uniformly quantized separately from other bands. (For example, 8-bit uniform quantization) In the above explanation, image signals are not specified in particular, but include multivalued black and white images, RGB color component images, luminance of Y, R-Y, B-Y, etc. , all dye signals are included in this image signal.

Furthermore, it can be applied to inter-frame difference signals in moving images such as television signals, and sufficient effects can be obtained. The interframe difference signal is described in detail in document (5). (Reference (5) “Television Bandwith Compression Transmission Bimotion Compression Interframe Coding”
Transmission by Motion−compensated Inter
-frame Coding)”, IE Communication Magazine (IEEE
Communication Magazine), November 1982 issue
(Pages 24-30) (Effects of the Invention) As described above, by using the image signal encoding method and encoding/decoding device of the present invention,
The orthogonal transform coefficients are divided into a plurality of categories according to the size distribution of the transform coefficients for each block, and different bands are realized for each category, making it possible to perform efficient vector quantization.

As a result, it is possible to reduce the occurrence of blurring in edge portions even for images with strong directionality, and to realize compression/expansion processing with high compression efficiency.

[Brief explanation of the drawing]

FIG. 1a is a block diagram showing an embodiment of the encoding device of the present invention, FIG. 1b is a block diagram showing an embodiment of the decoding device, and FIG. 2 is a diagram showing band division in the conventional system. , Figures 3a, b, and c are diagrams showing the functions of the image region and the transformation region, and Figures 4 a to d are diagrams showing examples of mask patterns used for category analysis.
FIGS. 5a, b, c, and d are diagrams showing an example of a band table. In the figure, 1...block reading unit, 2...
...DCT conversion section, 3...Category analysis section, 4...
...mask pattern storage section, 5...band division section,
6, 24...Band table storage section, 7...Vector quantizer, 8, 22...Codebook storage section, 9...Encoding section, 20...Decoding section, 21...
...Vector inverse quantizer, 23...Block synthesis section, 25...Inverse DCT transformation section, 26...Block writing section.

Claims (1)

[Scope of Claims] 1. On the transmitting side, an image signal is read out in units of blocks each consisting of a plurality of pixels, orthogonal transformation is performed on each block to obtain a plurality of transform coefficients, and the magnitude of the plurality of transform coefficients in the block is determined. A category number is generated by classifying each block into one of a plurality of categories according to the distribution of The transformation coefficients in the block are divided into multiple bands to generate multiple vectors, and the category number and
A plurality of vector indices are generated by vector quantizing the plurality of vectors according to a predetermined codebook for each category and each band, and a code is generated by encoding the category number and the plurality of vector indices. The receiving side inputs the code, decodes the category number and the plurality of vector indexes, and decodes the category number and the codebook predetermined for each category and band. The decoded vector indices are inverse vector quantized to generate a plurality of decoded vectors, and the decoded category number and the band division table in a predetermined block that is different for each category are used. Therefore, the plurality of decoded vectors are arranged to synthesize a block consisting of a plurality of decoded transform coefficients, and the plurality of decoded transform coefficients in the block are inversely orthogonally transformed to generate a decoded image signal of the block,
An image signal encoding/decoding method for outputting the decoded image signal in block units. 2. Means for reading out image signals in units of blocks consisting of a plurality of pixels, means for performing orthogonal transformation on the block units to obtain a plurality of transform coefficients, and means for obtaining a plurality of transform coefficients according to the size distribution of the plurality of transform coefficients in the block. means for classifying each block into one of a plurality of categories and generating a category number; means for dividing a transform coefficient into a plurality of bands to generate a plurality of vectors; and a means for vector quantizing the plurality of vectors according to the category number and a predetermined codebook for each category and each band. 1. An image signal encoding device comprising: means for generating a vector index; and means for encoding the category number and the plurality of vector indices. 3 Read out the image signal in units of blocks consisting of a plurality of pixels, perform orthogonal transformation on the blocks to obtain a plurality of transformation coefficients, and read out the image signals in units of the blocks according to the distribution of the magnitudes of the plurality of transformation coefficients in the blocks. A category number is generated by classifying into one of multiple categories, and the conversion coefficients within the block are divided into multiple bands according to the category number and a predetermined band division table within the block that is different for each category. generating a plurality of vectors, vector quantizing the plurality of vectors according to the category number and a predetermined codebook for each category and each band, and generating a plurality of vector indices; means for inputting a code from an image signal encoding device that encodes the category number and the plurality of vector indices to generate a code, and decoding the category number and the plurality of vector indices from the code;
means for generating a plurality of decoded vectors by inverse vector quantizing the plurality of decoded vector indices according to the decoded category number and the codebook predetermined for each category and each band; means for arranging the plurality of decoded vectors according to the decoded category number and the band division table in a predetermined block different for each category to synthesize a block consisting of a plurality of decoded transform coefficients; , means for generating a decoded image signal of the block by performing inverse orthogonal transform on a plurality of decoded transform coefficients in the block, and means for outputting the decoded image signal in units of blocks. Device.