JP2943925B2 - Image coding method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image encoding method that divides image information into a plurality of blocks each including a plurality of pixels, and performs encoding in block units.

[Conventional technology]

Conventionally, since image data has an enormous data amount compared to other encoded symbol data, various data compression / encoding methods and transmission methods suitable for storage / storage and data transmission have been proposed. ing.

As an encoding method, it has been proposed to divide image information into a plurality of blocks consisting of a plurality of pixels and encode them in units of blocks in order to enable high-rate compression of a halftone image or a multi-valued image.

In particular, from the viewpoint of data storage / storage, an image coding method that vector-quantizes an image into blocks of a plurality of pixels can achieve performance close to the rate / distortion limit in principle. Therefore, attention has been paid to its high compression ratio.

When performing normal vector quantization, a reproduction vector is designed by performing training using various images in advance. Then, at the time of encoding, a reproduction vector that minimizes distortion with respect to the input image is obtained, and this reproduction vector is encoded as a code. Therefore, how to optimally design the reproduction vector is an important point in vector quantization. Generally, in the design of a reproduction vector, an optimal solution of the reproduction vector is obtained by a vector quantization algorithm called an LBG method.

[Problems to be solved by the invention]

However, since this algorithm minimizes the average distance between the training image and the playback vector,
On the other hand, a reproduction vector biased for training a large number of patterns on average is obtained as an optimal solution.

In general, a vector quantizer is designed by performing training with various images having different frequencies, but a normal image still has many flat portions. However, the deterioration of the image in a high frequency portion such as an edge portion of the image is large. This is because the ratio of the edge part included in the training is small, and the amount of information in the edge part is larger than that in the flat part.

Another problem is that, when performing vector quantization, if the dimension of the vector increases, the amount of calculation required for quantizer design increases exponentially. Further, even if the reproduction vector can be designed, the amount of calculation for finding the vector with the minimum distortion for the input image also increases exponentially.

Furthermore, if hardware is to be realized, the scale of the hardware will be enormous. In particular, full search vector quantization becomes impossible.

Generally the vector quantization, the two-dimensional image and block of, and N-dimensional Euclidean space R ^N together N samples data is the configuration of the block, it said to be Matsupingu conversion of R ^N with respect to a finite number of Sabusetsuto. Therefore, as the number of sample images increases, the number of dimensions in the Euclidean space also increases, and hardware is one of the drawbacks.

In order to solve the above-described problems, a large number of edges are added to a training image, or training is artificially created.

However, when the reproduction vector is designed by such a method, the vector-quantized image becomes an unnatural image.

It is also conceivable to classify images according to their properties and design a reproduction vector for each classification, but this is not a fundamental solution. This is because the amount of information in the edge portion is larger than that in the flat portion (high-frequency portion) as described above. Further, these solutions cannot solve the second problem, that is, the problem of the amount of calculation and the hardware.

[Means for solving the problem]

The present invention has been made in view of the above points, and divides image information into a plurality of blocks each including a plurality of pixels.
An image coding method for performing coding in block units, wherein the image information is converted into a frequency space in block units to extract a plurality of transform coefficients, and the transform coefficients are divided into a plurality of bands in block units. Encoding the transformed coefficients, and further varying the number of bands in the division and varying the transform coefficients belonging to the same band to encode a code having a code length according to the characteristics of the image information. It is characterized by doing.

〔Example〕

The present invention will be described below using some preferred embodiments.

First, the principle of encoding used in the present invention will be described before the specific description of the embodiment.

First, this N-dimensional Euclidean space is divided into a plurality of sub-
Divide into Euclidean space and try to reduce the number of dimensions.

In this method, for example, the inside of a block is further divided into sub-blocks, and vector quantization is performed individually. Therefore, the number of constituent pixels of the sample data in the sub-block is N less than M, and the the R ^M Euclidean space. By setting M to a value that can be realistically implemented in hardware, the hardware of the vector quantizer becomes possible.

In order to facilitate division into sub-blocks, for example, block data is converted to a frequency space.
For this reason, orthogonal transformation to the frequency space is performed in block units.

Generally, when a frequency analysis is performed on an information source having a high correlation, such as an image, power tends to concentrate on low frequencies.
Utilizing such a property, an image is subjected to an orthogonal transformation, and then the transform coefficients are analyzed, so that the image is sub-block-divided according to the property, that is, the frequency, and each is independently vector-quantized. Since the division into the sub-blocks is performed in the frequency space, this is called band division. This eliminates a huge amount of calculation and hardware.

Second, after one block is divided into bands, if the block has a strong low-frequency component, only the low-frequency band is transmitted. This is the flat part (low frequency part) of the image
This is because the high frequency component need not be transmitted and only the low frequency band needs to be transmitted. This is called classification. Such classification is determined based on information of only the components in the block. This solves the problem of image quality degradation at the edge of an image.

Third, a transmission scheme is adopted in which the coded data classified into such bands and classes is transmitted in frame units from the low-frequency portion to the high-frequency portion. That is, the coded data in the low-frequency part of each block is transmitted for one frame,
Subsequently, the higher-frequency encoded data is transmitted for one frame, and further, the higher-frequency encoded data is sequentially transmitted for one frame, and so on. The resolution improves sequentially from low frequency to high frequency. Such a transmission method is called a progressive transmission method, and a transmission response with a coarse image is fast, and transmission can be interrupted in the middle. Such progressive transmission is performed in units of bands obtained by orthogonally transforming an image, and various operations are executed for each band described later.

In such progressive transmission, much power is concentrated on the flat part (low-frequency part) of the above-mentioned image, so that the outline of the image can be seen only by the low-band transmission, and the low-band code By using a shorter one, high-speed transmission of the low band itself is also possible.

Fourth, in the progressive transmission for each band, various conversions are performed in the real space or the frequency space for each band. This conversion is, for example, processing such as frequency correction (MTF correction) of an image, edge portion extraction, and affine conversion etc.

Hereinafter, an embodiment of the present invention using such an encoded transmission system will be described in detail.

FIG. 1 shows a schematic configuration of an image encoding apparatus to which the present invention is applied.

Reference numeral 1 denotes an orthogonal transform unit which performs orthogonal transform on an image input from the signal line 10 for each block (in this embodiment, a unit block is 4 × 4 pixels). In this embodiment, the orthogonal transform is 4
A Hadamard transform of × 4 pixels is used. By this Hadamard transform, each block image having 4 × 4 pixels as a unit block is converted into 16 conversion coefficients (Y ₁₁ , Y ₁₂ ,..., Y ₄₄ ). Reference numeral 11 denotes a signal line for these conversion coefficients. FIG. 2 is a conceptual diagram of the Hadamard transform. 4 × 4 in FIG. 2 (a)
The pixel data X ₁₁ , X ₁₂ ,..., X ₄₄ are subjected to Hadamard transform, and the second
Conversion to the sequence components of Y ₁₁ , Y ₁₂ ,..., Y ₄₄ in FIG. Y _{11 to} Y ₄₄ are components corresponding to the sequence at the same position in FIG. As can be seen from the figure, a higher i, j value of Y _ij indicates a higher frequency component.

Each component Y ₁₁ after orthogonal transformation output from the orthogonal transformation unit 1
The to Y _44, bands divided as in Figure 4. FIG. 4 (a)
4 shows an example of 4-band division, and FIG. 4B shows an example of 3-band division. Taking the three-band division of FIG. 4 (b) as an example, Y ₁₁ , Y
The three components of ₁₂ and Y ₂₁ are the first as band 1 (low frequency band).
Of the sub-block. Y ₁₃ , Y ₁₄ , Y ₂₂ , Y ₂₃ , Y ₃₁ , Y
_32, 7 components of Y ₄₁ constitute a second, third sub-block as well as a band 2 (intermediate frequency band), the other component band 3 (high frequency band).

Reference numeral 4 in FIG. 1 denotes a vector quantizer, which performs vector quantization on the output from the orthogonal transform unit 1 for each of the three sub-blocks described above. The vector quantizers 4a to 4c for the three bands are configured completely independently. By dividing the vector quantizer into these three bands,
The number of dimensions is reduced, and the scale of the quantizer can be reduced.

The vector quantizers 4a to 4c are constituted by a look-up table constituted by a ROM or the like so as to select an optimum reproduction vector when an input vector is input. Here, it is assumed that the optimum reproduction vector has been obtained from the training data divided into each band in advance.

Reference numeral 2 in FIG. 1 denotes a class classification unit for analyzing Y ₁₁ to Y ₄₄ from the orthogonal transformation unit 1 and classifying each block according to frequency.
The class classification unit 2 classifies the blocks into four classes for each frequency and outputs the result to the signal line 12 as classification information composed of two bits of data.

FIG. 5 shows a classification method in the classification unit 2.

 Here, it is defined as follows.

These E1, E2 and E3 are average values for each band shown in FIG. 4 (b).

As shown in FIG. 5, if E1.gtoreq.T1 (T1 is a threshold), the signal is classified into class 1 (flat portion) as a low-frequency class. Furthermore, if E2≥T2, it is class2 (medium edge) as a medium frequency class, and if E3≥T3, it is class3 as a high frequency class
(Large edge) E1 ≧ T1, E2 ≧ T2 and E3
Those that do not satisfy ≧ T3 have lower power of all bands
Classify as a flat part of Class1.

When such a classification is made, the class in which the power of the high frequency component is large, that is, the high frequency class is set so that the number of data bands to be transmitted is increased.

That is, as shown in FIG. 6, only the information of band 1 having a data length of 1 bit (data below the low-frequency component vector quantum) for class 1 (flat portion) and class 2 (data length of 1 for medium edge). Information of band 1 of bit and information of band 2 of data length m (middle frequency component), and information of band 1, band 2, and band 3 of data length l, m, n for class 3 (high frequency Component).

In this way, a high compression rate can be expected by classifying and band-dividing each block after the orthogonal transformation and adaptively encoding according to the image quality. This is because normal image information has many flat portions of class 1 represented by a short data length. As shown in FIG. 6, a class code is added to the head of each class of data as an index. When classifying into three classes, a 2-bit class code is required. Class1: Class code = 00 (binary display) class2: 〃 = 01 (〃) class3: 〃 = 10 (〃) You.

In some of the embodiments described later in this specification, the class code of class3 may be represented as “11” for convenience.
The length of the class code changes correspondingly according to the number of classifications.

Reference numeral 6 in FIG. 1 denotes a sequencer. Next, the function of the sequencer 6 will be described.

The reproduction vectors 5a to 5c of the respective bands obtained by the vector quantizers 4a to 4c are grouped for each band by the sequencer 6, and transmitted for each band based on the classification information 12.

Figure 7, for each block of the two-dimensional image data and block divided into B ₀₀ .about.B _mn shows the results of the classification of the above. (A) shows the number of each block, and (b) shows the result of classifying the block.

FIG. 8 shows a procedure for transmitting data obtained by encoding the image shown in FIG. 7 (a). In STEP1 constituting the one screen to transmit information of the band 1 of the total block B ₀₀ .about.B _mm. Then, information of band 2 is transmitted in STEP2, and information of band 3 is transmitted in STEP3.

In SETPE2 and SETPE3, it is not necessary to send the band information of all the blocks constituting one screen. For this reason, information of band 2 for blocks of class 2 and class 3 and information of band 3 for blocks of class 3 are transmitted only to blocks at necessary places.

For each such block, a band from steps 2 and 3
Information on whether or not transmission of band 2 information is necessary
At the time of transmitting the information of band 1 in the above, the class code is added to the information of band 1 of each block and transmitted for the total number of blocks. Then, this class code is stored for each block on the receiving side, and the information of band 2 and the information of band 3 are picked up only at a portion (block) where high band information is required by referring to the class code.

FIG. 9 illustrates a signal processing method on the receiving side. As the reception signal 20, information for each band of each block is input in the order of STEP shown in FIG.

When the class code is transmitted together with the band 1 information in STEP 1, the class code is stored in the class code memory.
It is temporarily stored in 21. The received code 20 is converted to real space data by the inverse Hadamard converter 22 and the real data output
Output as 23. The actual data output 23 is supplied to, for example, an image memory 24 or the like. An image based on the image data stored in the memory 24 is displayed on the display 25.

In STEPs 2 and 3, the processing for the received signal 20 differs according to the contents of the class code memory 21. The contents to be processed according to the contents of the class code are as follows: When the class code = 1 (00 in binary), there is no high band component of the block, and therefore, it is skipped.

When class code = 2 (01 in binary) Only the information of band 2 is subjected to inverse Hadamard transform.

When the class code is 3 (10 in binary), the information of the band 3 and the band 2 is subjected to the inverse Hadamard transformation. Therefore, the sequentially transmitted high band codes (the information of the band 1 and the band 3) are It is recognized as the target block information by the difference with the contents.

As described above, the information transmitted for each band is inversely transformed independently for each band, and is added to the previous inverse transformation value already stored in the memory 24 for each block. As a result, a final decoded image can be obtained.

As described above, the image data is divided into low-frequency components, medium-frequency components, and high-frequency components by using orthogonal transform, and each band is transmitted independently.
This has the effect of enabling transmission with good compression efficiency and good image transmission response.

The good image transmission response means that the whole image of the image is transmitted faster, though the gradation and the sharpness are coarse. By superimposing the data transmitted thereafter, the gradation and sharpness of the image are improved, but the entire image can be constructed as quickly as possible, and depending on the situation, even if the transmission is terminated halfway, It is possible to obtain an image of a reasonable image quality.

Such a response speed is particularly suitable for moving image transmission. In addition, even for a still image, it is extremely important that the entire image be seen quickly for image search because it is common to search for images visually.

Also, high compression efficiency reduces transmission time,
In addition, the capacity required for storage and accumulation is reduced, and the cost is great.

As described above, in the process of transmitting information for each band,
Various conversions are possible. Here we emphasize the high frequency components
An example will be described in which MTF correction is performed.

In general, as the transfer characteristics of the system, the signal components on the high frequency side become dull due to the transfer characteristics of the input device, analog processing system, etc.
Images often lose sharpness. For this reason, it is necessary to slightly increase the gain of the frequency component in the high frequency range on the transmitting or receiving side to correct it.

FIG. 10 shows data transmission formation when various conversions are required. Before transmitting information of each band, coefficients (parameters 1, 2, and 3) for correction of information of each band are transmitted in advance. . The parameters 1, 2, and 3 may not be transmitted for each band, but all coefficients of each band may be transmitted in advance. In the example of FIG. 10, it is assumed that the data is sequentially transmitted for each band.

Such a parameter gives a coefficient corresponding to the gain for the information of each band. For example, the gain in band 1 is α
₁ , the gain in band 2 is α ₂ , and the gain in band 3 is α ₃ , the Hadamard transform value Y ′ of the restored signal is Y ′ = α ₁ · (band 1 value) + α ₂ · (band 2 ) + Α ₃ · (value of band 3). By setting α ₁ <1, α ₂ > 1, α ₃ > 1, the high frequency range of the restored image can be raised.

Conversely, low-frequency emphasis can be performed by setting α ₁ > 1, α ₂ , α ₃ <1.

In this conversion operation, the gain (α _i ) * (the value of band i) is transmitted for each band. This is because the addition for each band is established (because of the linear combination), and the Hadamard transform (inverse transformation) is a linear transformation, so the inverse Hadamard transformation is performed for each band,
Real data can be superimposed.

Another method of transmitting the encoded data by the sequencer 6 is shown in FIG. Classification and band classification are 7th
The operation is performed on the image of FIG.
That is, in STEP 1, all the class codes of each block constituting one screen are all transmitted. In STEP2, the information of band 1 is used, in STEP3, the information of band 2 and in STEP4, the information of band 3
To transmit information.

In STEPs 3 and 4, it is not necessary to send the band information of all the blocks constituting one screen. Therefore, the information of band 2 and band 3 is transmitted only to the necessary block.

Since information on whether or not transmission of information on bands 2 and 3 is necessary in steps 3 and 4 is performed by transmitting the class codes for all blocks at the time of STEP 1, this class code is stored on the receiving side. The information of the band 2 and the band 3 is picked up only at the portion (block) where the high band information is required by referring to the information.

On the receiving side, the data in STEP1
It is stored in the class code memory 21 shown in the figure, and in the case of STEP 2, all blocks are sequentially processed by the inverse Hadamard converter 22, and STEP 3 and STEP 4 are exactly the same as STEP 2 and STEP 3 in the above-described embodiment. .

In the case of this transmission method, since image data is not transferred until the transmission of the data in STEP 1 is completed, a slight delay in image transmission response occurs, but the process of separating the class code and band data in STEP 1 is not necessary. The load on the demultiplexer can be reduced.

The data conversion in the data transmission procedure shown in FIG. 11 will be described.

That is, as shown in FIG. 12, following the transmission of the class code in STEP1, the parameters 1, 2, and 3 representing the correction coefficient are transmitted.

Here, the orthogonally transformed data is subjected to an inverse transform on the receiving side, converted again to the original actual data, and then a process of performing a spatial filter operation is performed for each band (or for each ST in FIG. 10 (B)).
An example of performing this for each EP will be described.

 FIG. 13 shows a flowchart of the conversion process.

Reference numeral 25 denotes the transmission of STEP 1 shown in FIG. This parameter is a parameter for performing conversion. For example, in the case of contour enhancement, selection of a Laplacian filter for edge detection of 5 × 5 or 3 × 3 shown in FIGS. In the convolution calculation (convolution calculation) at the time of processing, the setting of a coefficient β in [output data = center pixel data + β · (Laplacian output)] is designated.

26 indicates transmission of each band data in STEP2 to STEP4. Inverse Hadamard transform is performed on the received band data at 27. The result at this time is to be added to the image memory. Blocks that do not require band data from the class code information are skipped (SKIPed).

A spatial filtering operation is added 28 to the data that has been inversely transformed and returned in this manner. This is done according to the parameters described above. In some cases, this parameter may be set on the receiving side.

The above operation is repeated by the number of bands. In this way, conversion for MTF correction can be performed.

Still another method of transmitting encoded data by the sequencer 6 will be described.

That is, the processing of STEP 2 in FIG. 11 is performed first,
Next, STEP1, STEP3, and STEP4 are performed.

As a result, after transmitting all the data of the low sequence component (band 1), the class code and the data of the middle sequence component and the high sequence component are sequentially transmitted. As a result, compared to the case where the transmission procedure shown in FIG. 11 is performed, the image response is further accelerated, and an image represented by only low-sequence component data can be obtained more quickly.

In the data transmission procedure shown in FIG. 8 or 11, it is expected that more efficient compression efficiency can be obtained by transmitting the class code with run-length coding. In this case, an encoder and a decoder for the run-length encoding are required in the encoder and the decoder, respectively.

Still another method of transmitting encoded data by the sequencer 6 will be described.

In this embodiment, classification and band division are performed in exactly the same manner as in the above-described embodiment. However, the class code 3 having the information of the band 2 and the band 3 is represented by “11”, and the data transmission procedure is performed according to the procedure shown in FIG.

That is, in STEP 1, data of band 1 of all blocks constituting an image is transmitted. In step 2, bit data representing whether or not all the blocks constituting the image each have band 2 data are expressed by 1 bit. 0
Means that there is no band 2 and 1 means that there is band 2. Next, the data of band 2 is transmitted in STEP3. In STEP 4, bit data expressing whether or not all the blocks constituting the image each have band 3 data is represented by 1 bit. 0 means no band 3 and 1 means band 3. Next, the data of band 3 is transmitted in STEP3.

On the receiving side, the data in STEP 1 is sequentially decoded as it is, in STEP 2 the data is stored in the class code memory 21 in FIG. 9, and in STEP 3 based on the data held in the class code memory 21, The data of band 2 is decoded as the value of the corresponding block. In STEP4, the data is stored in the class code memory 21 of FIG. 9, and in STEP5, based on the data held in the class code memory 21,
The data of band 3 is decoded as the value of the corresponding block.

In this embodiment, although the number of transmission STEPs increases, the ninth
The capacity of the class code memory 21 shown in the figure can be reduced as compared with the above-described embodiment.

Further, in STEP 4 in the embodiment of FIG. 15, band 3 is transmitted because part or all of the block to which band 2 is transmitted, and as shown in FIG.
0 or 1 is sent to only the part where the data of EP2 is 1. That is, information is transmitted on whether or not the band 3 is further added. Although the processing in STEP 4 is slightly more complicated than that in FIG. 15, the compression efficiency is further improved.

 FIG. 17 shows still another transmission procedure of the encoded data.

Classification and banding are performed in exactly the same manner as in the procedure shown in FIG. However, the data transmission procedure is performed as shown in FIG.

That is, in STEP 1, all the information of band 1 of all the blocks in the image is transmitted. In STEP 2, whether each block in the image has band 2 or not is represented by 1 bit, meaning that 0 does not have 1 means that 1 has. If so, the band immediately follows "1". 2 is transmitted in a form to which data is added. In STEP3, data of band 3 is transmitted as in STEP2.

On the receiving side, in STEPs 2 and 3, this block is band 2
The position of the data having the band 2 can be easily determined by looking at whether the data has the data of the band 3 or not, and the data of the band 2 and the data of the band 3 can be sequentially decoded.

In the transmission procedure of FIG. 18, if a block in which no band 2 or band 3 exists in STEP 2 and STEP 3, a bit of 0 will continue for a long time. The continuous string of 0 bits is run-length encoded as shown in FIG. As a result, more efficient compression can be expected as compared with the procedure in FIG.

Next, an embodiment configuration in which sub-blocks for band division and class division are different will be described. It is assumed that the band division is the same as in the above-described embodiment.

Regarding the classification, in order to analyze and classify using the Hadamard transform coefficient, FIGS. 9 (a), (b),
Three edge quantities as shown in (c) are defined. Each element in the block of FIG. 19 corresponds to FIG. 2 (b). (A) is a calculation example for obtaining the edge amount ED. When Hadamard transform is performed on an image with edges, power concentrates on this part, and ED = | Y
₁₂ | + | Y ₁₃ | + | Y ₂₁ | + | Y ₂₂ | + | Y ₃₁ | Therefore, utilizing this property, the ED
If T1 (T1 is threshold), classify as class1 (flat part). In the case where the image has a vertical edge, the power is concentrated on the portion (b), and VE = | Y ₁₂ | + | Y ₁₃ | +
| Y ₁₄ | When the edge amount VE is large and the image has a horizontal edge, the power is concentrated on the portion (c),
HE = | Y ₂₁ | + | Y ₃₁ | + | Y ₄₁ | In addition, if the difference | VE−HE | between the vertical edge amount VE and the horizontal edge amount HE is small, it is often a slanted edge, and as shown in FIG. 20, class2 (vertical edge) class3 ( Horizontal Etch) class4
Classify as (named edge). In the figure, T2 is a threshold for determining a slanted edge.

FIG. 21 shows a code form based on each classification. class1
The coded data for the flat part is only the information of band 1, and the coded data for the edge parts of class 2 to 4 is composed of all the information of band 1 to band 3. This is because the vertical edge, the horizontal edge, and the slanted edge all have a high edge component frequency.

The class code is 2 bits to represent 4 classes. This is because the reproduction vector by the vector quantizer of each band is different for each class.
This also changes the full search vector quantization described above to a partial search type.

FIG. 22 shows a configuration example of an image encoding device. That is, the 22nd
In the figure, a classifying output signal 53 indicating the above-described classifying is input to a vector quantizer 54 for coding each block output from the orthogonal transform unit 51. The table space of the look-up table of the vector quantizer 54 is switched by the input signal for classification, and a reproduction vector suitable for each class can be obtained.

Similar to the sequencer 6 in the above-described embodiment, the processing by the sequencer 56 transmits encoded data in accordance with the classification signal from the classification unit 52 in accordance with the procedure shown in FIG. Also. The orthogonal transformation unit 51 performs Hadamard transformation for each block in the same manner as in the above-described embodiment, and the vector quantizer 54 performs vector quantization for each band using the quantizers 54a to 54c for each band. Done.

Next, an example in which band division is adaptively determined according to each classification will be described. Figures 19 and 20 for classification
It is the same as that shown in the figure.

In FIG. 23, 63 is a band dividing unit. A 2-bit signal indicating the four classes classified by the class classification unit 62 is a signal line.
The signal is input to the band dividing unit 63 via 72. Flat portion as in the Y ₁₁ to Y ₄₄ in response to the band dividing section 63 classification signal Figure 24 (class1), the vertical edge portion (class2), lateral edge portions (class3), oblique edge portion (class4 ) To split the band. Figure DC is DC component of the image in Y _11, 1, 2 each band number, the hatched portion is masked without quantization (0
) Part.

This is because the hatched portion has a very small power. Generally, the flat portion of an image has a smaller amount of information than the edge portion. In this embodiment, utilizing this point, the flat part (class1) is divided into DC and band 1 only, and the edge part (class2, 3, 4) is divided into DC and band 1 and band 2. This eliminates the redundancy of the flat part. That is, an image having a higher frequency is provided with more bands. In addition, the dimension of each band is set to 7,
By limiting to eight dimensions, the scale of vector quantization is reduced and full search is possible.

If the number of coding bits to be given to each class is, for example, as shown in FIG. 25, a large amount of information can be given to the edge portion and the coding compression rate can be increased.

In FIG. 25, class 1 is an example in which 6 bits are assigned to DC, 8 bits are assigned to band 1, and classes 2 to 4 are assigned 6 bits to DC, and 8 bits are assigned to band 1 and band 2, respectively. Each of the class bands 1 and 2 is independently vector-quantized by the vector quantizers 64 and 65, respectively.

In FIG. 23, a signal line 73 is a signal line for sending the DC component (6 bits) divided by the band dividing section 63 to the encoder 66. 74,75
Respectively calculate the Hadamard transform coefficients of the band 1 and the band 2 divided by the band dividing unit 63 by the vector quantizer 64 and
This is the signal line to send to 65. Look-up tables are written in the vector quantizers 64 and 65 so as to select an optimum reproduction vector when an input vector (band 1 and band 2) is input. Here, the same class classification and band classification processing as described above are performed in advance on images having various frequencies, and training is independently performed on bands 1 and 2 to obtain an optimal reproduction vector, thereby forming a lookup table. Both band 1 and band 2 seek 256 types (8 bits) of reproduction vectors.
However, for class1, only vector quantization of band 1 is performed. Further, the quantizers 64 and 65 are connected via a signal line 72,
A 2-bit classification signal is input, and the contents of the lookup table are selected for each class.

In the encoder 66, the class classification signal 72, the DC component 73, the band 1
And the vector quantization results 76 and 77 for band 2 are coded and coded into the format shown in FIG. The coded image can be sent to a memory or a communication line via a signal line 78.

In the embodiment described above, the Hadamard transform is used as the frequency analysis. However, the present invention can be implemented by using another orthogonal transform such as a cosine variable or a KL transform. In the vector quantizer, a look-up table is used. However, the vector quantizer may be constituted by a microprocessor to calculate an optimum reproduction vector.

Also, the size and shape of the coding unit block are not limited to those of the present embodiment, but are appropriately selected according to the image density to be coded, and the like. It goes without saying that it is set to a number.

(Effect)

As described above, according to the present invention, there is provided an image coding method for dividing image information into a plurality of blocks each including a plurality of pixels, and performing coding in block units, wherein the image information is converted into a frequency space in block units. Extracting a plurality of transform coefficients, dividing the transform coefficients into a plurality of bands in block units, encoding the divided transform coefficients,
By making the number of bands in the division variable and making the transform coefficients belonging to the same band variable and encoding them into a code having a code length corresponding to the characteristics of the image information, according to the characteristics of the image for each block Band splitting according to the importance of frequency components can be performed flexibly, and adaptively band-divided blocks can be further subjected to variable-length coding according to the characteristics of images, resulting in efficient transmission. It can be performed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration example of an encoding apparatus to which the present invention is applied, FIG. 2 is a conceptual diagram of Hadamard transform, FIG. 3 is a diagram showing sequence components, and FIG. 4 is a diagram showing an example of band division. FIG. 5 is a diagram showing a classification process, FIG. 6 is a diagram showing encoded data, FIG. 7 is a diagram showing an example of image classification, and FIG. 8 is a diagram showing a transmission format of the encoded data. , FIG. 9 is a diagram showing an example of a processing block on the receiving side, FIGS. 10 and 12 are diagrams showing a transmission format of coded data to which parameters are added, FIG. 11, FIG. 15, FIG. 17, 18, 21 and 25 show other transmission formats of the encoded data, FIG. 13 shows a procedure of the conversion process, and FIG. 14 shows an example of the Laplacian filter FIG. 19, FIG. 19 and FIG. 24 are diagrams showing another classification method, FIG. 20 is a diagram showing another classification procedure, FIG. And FIG. 23 is a block diagram showing another configuration example of the encoding device. 1 is an orthogonal transformation unit, 2 is a class classification unit, 4 is a vector quantizer, 6 is a sequencer, 21 is a class code memory, 22 is an inverse Hadamard transformation unit, and 24 is a memory.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Yoshihiro Ishida 3-30-2 Shimomaruko, Ota-ku, Tokyo Inside Canon Inc. (72) Koji Hirabayashi 3-30-2 Shimomaruko, Ota-ku, Tokyo Canon (72) Inventor Mitsuru Maeda 3-30-2 Shimomaruko, Ota-ku, Tokyo Canon Inc. (56) References JP-A-63-215281 (JP, A) JP-A-63-1117583 (JP) JP-A-64-1383 (JP, A) JP-A-63-268387 (JP, A)

Claims (1)

(57) [Claims] 1. An image coding method for dividing image information into a plurality of blocks each including a plurality of pixels and performing coding in block units, wherein the image information is converted into a frequency space in block units to obtain a plurality of transform coefficients. Extracting, dividing the transform coefficient into a plurality of bands in block units, encoding the divided transform coefficients, and further changing the number of bands in the division and changing the transform coefficients belonging to the same band. Wherein the image information is encoded into a code having a code length corresponding to the characteristics of the image information.