JP3235510B2 - Encoding method and encoding device, decoding method and decoding device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an encoding method and apparatus for encoding a data string in which image data and its attribute data are mixed, a data string obtained by expanding a multi-valued image into a level plane, and the like, and an encoding method therefor. The present invention relates to a decoding method for decoding encoded data encoded by such an encoding method and apparatus, and an apparatus therefor.

[0002]

2. Description of the Related Art As one of coding techniques for a binary image, there is a prediction coding method for predicting a state of a target pixel from peripheral pixels and coding a result of the prediction. In predictive coding, surrounding pixels considered to have a correlation with a target pixel are used for prediction as a Markov information source.

As the prediction method, there are known a method of directly determining a predicted value and a method of indirectly determining a predicted value. As a method of directly determining a predicted value, for example, Japanese Patent Laid-Open No. 5-31637
As disclosed in Japanese Patent Laid-Open No. 0, there is a method in which a predictor with a high hit ratio is adaptively switched from a plurality of predictors, and one predictor is finally selected and used for prediction of an original image. . Also, as a method of indirectly determining the predicted value, for example,
There is a method in which information such as a predicted value is stored in a memory, and information in the memory is read from peripheral pixel values and used for prediction.

A prediction method disclosed in Japanese Patent Publication No. 6-95645 is a latter prediction method in which a prediction value is indirectly determined using a memory. The prediction error data is generated using the prediction value output in the up-up method. However, all the positions of the predicted pixels are not fixed, but the position of only one pixel is made variable, and the predicted position is changed by the predictive accuracy. ITU-TT. 82, JBIG (Joint Bi-level) defined by T-85
Image coding expert Grou
The prediction method used in p) is the above-mentioned Japanese Patent Publication No. 6-956.
In the prediction method disclosed in Japanese Patent Publication No. 45, a learning type method in which information in a prediction memory is updated according to an encoding state is adopted. For encoding, prediction error data that is a result of predicting a target pixel by the above-described prediction method is used.

[0005] In JBIG, arithmetic coding is employed in a coding unit for coding prediction error data. In arithmetic coding, for example, as shown in Japanese Patent Publication No. 1-17295, a region on a number line is divided according to the probability of occurrence of 1,0 of an input symbol, and the region is divided by the value of an input symbol. Select At this time, a symbol (dominant symbol) that matches the prediction is assigned to MPS (More Probe).
Able Symbol), a symbol that does not match (recessive symbol) is converted to an LPS (Less Probable S
ymbol). The estimated value of occurrence probability of MPS and LPS is changed by predictive predictive value or the like. JBIG
In the above, the symbol appearance probability estimating unit has a probability value (= region width LSZ) at the time of LPS occurrence and a learning function of updating the probability estimation value in accordance with the symbol appearance process. The learning function of the probability estimation is linked with the learning function of the predicted value described above. This division selection process is performed for the output sequence of the information source, and the output sequence is represented by a binary decimal representation of one point in the area obtained by the final division.

FIG. 13 is a conceptual diagram of arithmetic coding. The shaded areas are the effective areas after the symbol encoding is completed. In the initial state, the effective area is [0, 1).
This is a section, and the area width LSZ (0) at the time of LPS occurrence is set. 1-LSZ (0) is the area width when MPS occurs. Here, when the symbol is “0”, MPS,
LPS is defined as “1”. In FIG. 13, the probability values are represented by binary numbers.

[0007] In FIG. 13, since the first input symbol is “0”, the effective area is changed as shown in FIG.
3 (A) is a hatched area. Then, a new area width LSZ (1) for dividing the effective area into areas at the time of LPS generation and at the time of MPS generation is set. When the next symbol is "1", the occurrence of LPS causes
In (B), the hatched area is the effective area. Similarly, when the third symbol is “0”, the hatched area in FIG. 13C becomes an effective area, and when the fourth symbol is “1”, FIG.
The area shaded in 3 (D) becomes the effective area.

The output code is represented by C. The code for the input symbol sequence “0101” is shown in FIG.
What is necessary is just one point in the effective area in (D). Here, the base of the effective area, that is, the smallest value C (3) of the effective area is represented as a code and output.

In arithmetic coding, during the coding process,
Since one point in the effective area is used as a code, if the area is made large, the effective digits are reduced and the code amount is reduced. That is, by reducing the value of LSZ as much as possible when MPS occurs and increasing the value of LSZ as much as possible when LPS occurs, it is possible to secure a large effective area and reduce the code amount.

FIG. 14 is a schematic diagram of a predictive coding system employed in JBIG. In the figure, 1 is image data, 2
Is a context generation unit, 3 is a context, 4 is a prediction unit, 5 is a state, 6 is a probability estimation table, 7 is LSZ
SW and 8 are prediction values, 9 is an arithmetic coding unit, 10 is code data, 11 is rewrite data, and 12 is a rewrite signal.

The image data 1 is input to an arithmetic coding unit 9 for coding, and is also input to a context generation unit 2. The context generation unit 2 creates a prediction model template using neighboring pixels in order to predict a target pixel. FIG. 15 is an explanatory diagram of an example of a template used in JBIG. Assuming that P is the target pixel, a template using the reference pixels X _{9 to} X ₀ shown in FIG. 15 is generated. The reference pixels X _{9 to} X ₉
A value of ₀ becomes a 10-fold Markov information source of 2 ¹⁰ states (contexts), and is directly input to the prediction unit 4 as a context 3 and used for prediction.

The prediction unit 4 stores a prediction table in which a prediction value 8 corresponding to the values of the reference pixels X _{9 to} X ₀ and a state 5 for referring to the probability estimation table 6 are paired. The context 3 that is the value of the reference pixels X _{9 to} X ₀ is directly input as an address into the prediction table, and one of the 1024 entries in the prediction table is selected. The predicted value 8 and the state 5 are extracted from the entry of the prediction table selected by the context 3, the predicted value 8 is input to the arithmetic coding unit 9, and the state 5 is input to the probability estimation table 6. FIG. 16 is an explanatory diagram of an example of the prediction table stored in the prediction unit 4. FIG.
The prediction table shown in FIG. 6 stores a pair of a predicted value and a state corresponding to an address. Note that the address is indicated by a binary number, which corresponds to each pixel value of the reference pixels X _{9 to} X ₀ input as the context 3. For example, if only the value of the reference pixel X ₀ is “1” and the other reference pixels are “0”, the address “00000000001” is referred to, and the predicted value “1” and the state “2” are obtained.

The probability estimation table 6 corresponds to the state 5 and includes a region width LSZ when an LPS occurs, NLPS as a next state value when an LPS occurs, and NMPS as a next state value when an MPS occurs. SW, which is a flag indicating whether or not the predicted value is rewritten when the normalization occurs, is stored. The LSZ and SW7 corresponding to the input state 5 are output to the arithmetic coding unit 9, and NM
The PS and the NLPS are sent to the prediction unit 4 as a part of the rewrite data 11. FIG. 17 is an explanatory diagram of an example of the probability estimation table 6. In FIG. 17, LSZ, NMPS, NLPS, and SW are stored as a pair corresponding to the state 5 number. For example, when the state is “2”, the LSZ is “0x1114”, the NLPS is “16”, and the NMP
"3" is obtained as S and "0" is obtained as SW. Here, numerical values starting with “0x” are hexadecimal numbers.

The arithmetic coding unit 9 compares the prediction value 8 input from the prediction unit 4 with the image data 1, and if the values match, it is assumed that the symbol to be coded is "0", that is, that MPS has occurred. If they do not match, it is assumed that the symbol to be encoded is "1", that is, LPS has occurred. Using the LSZ input from the probability estimation table 6, FIG.
The arithmetic coding is performed as described in 3 and the code data 10 is output. When the LPS occurs, the new predicted value obtained by inverting the predicted value is rewritten only when the SW input from the probability estimation table 6 enables the rewriting of the predicted value (for example, when the SW is “1”). The data is sent to the prediction unit 4 as a part of the data 11 and the rewrite signal 12 is sent to rewrite the prediction value of the prediction table stored in the prediction unit 4.

In the JBIG, the arithmetic coding unit 9 is adjusted so that the effective area is always 0.5 or more and less than 1 so as not to increase the effective digits of the effective area.
This process is called normalization. The value of LSZ is always 0.5
The probability estimation table is configured so as to be a value less than or equal to. As a result, the effective area becomes less than 0.5 when LPS occurs, and normalization always occurs. Further, even when the MPS occurs, the new effective area is set to the LSZ from the current effective area.
May be less than 0.5 because it is obtained by subtracting
There are cases where normalization occurs. When this normalization occurs, a rewrite signal 12 is sent to the prediction unit 4. The prediction unit 4 writes the NMPS given from the probability estimation table 6 to the prediction table as the state of the context 3 given from the context generation unit 2 when the MPS occurs. When an LPS occurs, the NLPS provided from the probability estimation table 6 is written to the prediction table as the state of the context 3. After the prediction table is updated in this way, the processing for the next image data 1 is started.

A learning-type prediction / probability estimation method in which the above-described prediction unit 4 and probability estimation table 6 are sequentially rewritten will be further described using specific examples. FIG. 18 is an explanatory diagram of a specific example of input image data. FIG.
Now, the image data 1 of the template portion shown in FIG.
Only shows. As shown in FIG. 18, consider the case where the target pixel P is 1 when the contexts X _{9 to} X ₀ are “00000000011”. The prediction table and the probability estimation table 6 of the prediction unit 4 are shown in FIG.
It is assumed that the content shown in FIG. 7 is set.

First, the context "000000001
"1" is input to the prediction unit 4 as the context 3. At this time, the address “00000” of the prediction table of the prediction unit 4
For example, as shown in FIG. 16, the predicted value “0”
And the state value “80” are stored. The state value “80” read from the prediction table is input to the probability estimation table 6. Next, the state number “80” of the probability estimation table 6 is referred to. Since the contents shown in FIG. 17 are set in the probability estimation table 6, the LSZ value “0x5832” corresponding to the state number “80” is read and output to the arithmetic coding unit 9.

At this time, the value of the target pixel P is “1”, but the prediction value read from the prediction table of the prediction unit 4 is “0”, which means that LPS has occurred. When LPS has occurred, the area corresponding to the LSZ value read from the probability estimation table 6 is basically the next effective area. Since LSZ is less than 0.5, 0.5
Normalization is performed so as to be less than 1 and the NLPS value “80” of the state number “80” is overwritten as the state value of the address “00000000011” in the prediction table of the prediction unit 4. Further, since the SW of the state number “80” in the probability estimation table 6 is “1”, the predicted value is inverted to “1” and is overwritten as the predicted value of the address “00000000011” in the prediction table of the prediction unit 4. You. FIG. 19 is an explanatory diagram of a specific example of the prediction unit table after rewriting. In this way, as shown in FIG. 19, the prediction value and the state value of the address “00000000011” of the prediction unit are rewritten. The prediction of the next image data 1 is performed using the rewritten prediction unit. Next, when the same context is input, the predicted value is 1.

Although LPS has occurred in this example, when MPS has occurred, the prediction unit 4 performs normalization.
Is rewritten to NMPS “81”. When the same context 3 is input next after rewriting, the state value is “81”,
The SZ is “0x4d1c” from FIG. This value is
It is smaller than the LSZ “0x5832” when the state value is “80”, which means that the probability that an LPS will occur is small. This is because the prediction at the same context last time was hit, so that the next time the same context is input, the area width at the time of occurrence of MPS is widened and the effective area is widened when the prediction is hit again. Thus, it is possible to increase the final effective area and reduce the code amount.

As described above, the prediction unit is sequentially rewritten according to the probability estimation table, the prediction accuracy is improved, the number of MPS occurrences is increased, and the region width at the time of MPS / LPS occurrence is controlled to perform coding. Improving efficiency.

Also, the decoding process is performed while dividing the area, as in the encoding process. The difference from the encoding is that at the time of encoding, the prediction value is compared with the target pixel,
While the S / MPS area is calculated while being calculated, at the time of decoding, it is calculated by calculating which area of the LPS / MPS the code belongs to, and then the pixel of interest is determined from neighboring decoded reference pixels. Obtain while doing reverse prediction.

FIG. 20 is a schematic diagram of a predictive decoding system employed in JBIG. In the figure, the same parts as those in FIG. 14 are denoted by the same reference numerals, and description thereof will be omitted. 21 is an arithmetic decoding unit, 22 is an inverse prediction unit, 23 is an inverse prediction value, and 24 is rewritten data. The context generation unit 2 and the inverse prediction unit 22 operate exactly the same as the context generation unit 2 and the prediction unit 4 at the time of encoding. The inverse prediction unit 22 is the same as the prediction unit 4 at the time of encoding. As a matter of course, the same probability estimation table 6 as that used in encoding must be used.

First, as in the case of encoding, a context is generated from the values of reference pixels around the pixel of interest in the decoded image data 1, and the inverse prediction unit 22, the probability estimation table 6
, The LSZ and the inverse prediction value 23 are output to the arithmetic decoding unit 21.

Next, the code data 10 is supplied to the arithmetic decoding unit 21.
, After the comparison of the effective areas is performed, the inverse prediction value 2
3 is used to output image data 1. FIG. 21 is an explanatory diagram of the basic operation of the arithmetic decoding unit. First, the LSZ width input from the probability estimation table is subtracted from the current effective area to obtain (effective area−LSZ). The MPS area and the LPS area are divided based on this value. The next input code C is
A comparison is made as to whether the value is larger than the value of (effective area-LSZ) previously obtained. That is, C> effective area−LSZ is determined. As a result of this comparison, the pixel of interest is L
Since it can be determined whether the pixel is the PS or the MPS, the level of the target pixel is obtained from the input inverse prediction value. That is, in the case of MPS, the value of the pixel of interest is an inverse prediction value, and L
In the case of PS, the inverse prediction value may be inverted to be the value of the pixel of interest. With the above processing, the target pixel can be basically decoded.

However, in order to decode the next pixel, further processing such as determination of a new effective area and subtraction of a sign-effective area is required. For example, in the example shown in FIG. 21, since the code data C exists in the LPS area, the LPS area is set as a new effective area, a new LSZ is set, and the above processing is repeated. Further, similarly to the encoding, the prediction value and the state value of the prediction table stored in the prediction unit 4 are also rewritten.

With the above processing, highly efficient encoding of binary images and decoding of encoded data are achieved in JBIG. JBIG is a standard for binary images, but application to other image data is also being studied. For example, Hiroshi Yasuda, "International Standard for Multimedia Coding",
1991, Maruzen, p. 80-81 suggest applying an arithmetic coding technique to a multi-valued information source. Further, in order to apply arithmetic coding, which is a coding method of binary data, to a multi-valued information source, the multi-valued information source may be represented by a binary representation using, for example, a level-plane or bit-plane conversion method. It states that what should be done is.

FIG. 22 is an explanatory diagram of bit plane development. In a multi-valued image, the density is indicated by a plurality of bits per pixel. Here, the most significant bit of the bit string representing the density of each pixel is shown as MSB, and the least significant bit is shown as LSB. When developing a bit plane, it is assumed that one bit at the same position of each pixel is extracted and each plane is created. For example, an MSB plane is created by extracting only the MSB of each pixel. Each created plane is composed of 1-bit data for each pixel, and can be handled in the same way as a binary image.
Each plane can be encoded using the above-described JBIG.

Normally, it can be considered that there is not much change in the density between adjacent pixels of an image. Therefore, this method makes it easy to predict a high-order plane close to the MSB and reduces the code amount. However, it is known that, since the value changes in a low-order plane due to a small change in density, it is difficult to predict and the code amount increases. To solve this, Japanese Patent Laid-Open Publication No. 6-2
As disclosed in Japanese Patent Application Laid-Open No. 61214 and the like, a method of giving a different prediction model to each plane, a method of using a gray code conversion when developing a bit plane, and the like are also known. However, in any case, since codes are generated separately in each plane, it is necessary to manage codes for each of these planes. Also, at the time of decoding, since each pixel value cannot be determined unless all planes are decoded, a decoding circuit is required for each plane, or a memory for storing all images to be sequentially decoded is required. There is.

FIG. 23 is an explanatory diagram of the level plane development. In the case of level plane development, a plurality of bits representing the density of each pixel, that is, bits from the MSB to the LSB are directly connected. FIG. 23B shows each pixel of the multi-valued image by a rectangle, and the pixels marked with numbers are respectively referred to as pixel 0, pixel 1, pixel 2,. . . And As shown in FIG. 23A, in the data obtained by level plane development, bits from MSB to LSB of pixel 0 are arranged, and bits from MSB to LSB of pixel 1 continue.
Further, continuous data such as bits of pixel 2 is obtained.

In the case of using the learning type prediction method represented by the above-described JBIG, the prediction value and the state are rewritten so as to be close to the optimum state, so that there is an advantage that the predictive accuracy is improved. However, when heterogeneous data having a weak correlation is continuously input, if the prediction is performed according to the prediction method according to the JBIG standard, data with low correlation is used for the prediction. And the code amount may increase.

As described above, in the method of encoding data that has undergone level plane expansion, the encoding / decoding processing is one
Only one image is required, and code management is easy. However, the correlation between adjacent bits is weak in the data obtained by level plane expansion. For example, considering that encoding processing is performed using a template as shown in FIG. 15 where importance is placed on the correlation between bits, for example, each bit indicating the density in the same pixel or the high-order bit and the low-order bit of an adjacent pixel That is, bits having low correlation are used for prediction. For this reason, there is a problem that the accuracy of prediction is reduced and the prediction efficiency is reduced on average as compared with the case of bit plane development.

FIG. 24 is an explanatory diagram of an example of a data string developed on a level plane. In FIG. 24, the density of each pixel is indicated by 3 bits. For example, considering the case where the template shown in FIG. 15 is used, the first bit (M
SB) is taken as the pixel of interest. FIG.
The pixel corresponding to the reference pixels X ₀ and X ₁ is “0” in the third bit (LSB) of the pixel 0 and “1” in the second bit. The slight difference in the density of the pixel 0 does not affect the large change in the density of the pixel 1.
The bit is considered to have a weak correlation with the first bit of the pixel 1. Conversely, when the first bit of pixel 1 is set as the target pixel, the first bit of pixel 0, which is considered to have a strong correlation, is not used for prediction. As described above, since bits having low correlation are used for prediction without using bits having high correlation, there is a possibility that the predictive accuracy is reduced and the code amount is increased.

As shown in this example, 3 bits per pixel (=
In the case of an image of about 8 gradations, it can be considered that there is a certain degree of correlation between bits of each level.
When the number of bits is about 256 (= 256 gradations), it is easily presumed that the correlation between the low-level bits becomes weak and the prediction efficiency becomes poor. Furthermore, by using a model with a weak correlation in this way, a prediction prediction state is shifted to a state where the probability estimation state is worse than a steady state, even in a context in which the coding efficiency is expected to be increased, such as when the reference pixels are all 0s. There is also the problem of letting them do it.

On the other hand, it is conceivable to perform JBIG encoding on image data including attribute data. FIG. 25 is a block diagram showing an example of a case where intermediate information is stored with attribute data in a conventional image input / output device. In the figure, 101 is an input unit, 102 is an image processing unit, 103 is an image encoding unit, 104 is an image code data holding unit, 105 is an image decoding unit, 106 is a TI separation unit, 107 is a TI flag encoding unit, 108, a TI flag code data holding unit; 109, a TI flag decoding unit;
0 is an output unit. For example, in a copier or the like, when an image or the like input by a scanner unit is output to an output unit, image processing such as density adjustment and filter processing is generally performed on the input image. In order to improve the processing accuracy of these image processings, an input image is subjected to text image separation (hereinafter referred to as TI separation) or frequency band separation processing, and different image processing is performed according to each attribute. Have been. For example, when performing the TI separation, together with the original image data, attribute data indicating whether each pixel in the image is a text image portion or other, for example, a photographic image portion is added. In addition to the above, for example, in the case of a character document, attribute data indicating a contour portion of a character and a portion other than the contour, and in the case of a binary image, dither processing, halftone processing, error diffusion processing, etc. are separated. There are things.

In the configuration shown in FIG. 25, the image data input by the input unit 101 is transmitted to the input unit 1 by the image processing unit 102.
01 and other common image processing.
The image encoding unit 103 encodes the image data using a method such as JBIG, and stores the encoded data in the image encoded data holding unit 1.
04. On the other hand, the same image data input by the input unit 101 is subjected to TI separation processing in the TI separation unit 106, and a TI separation flag is generated for each pixel or a plurality of pixel groups. This TI separation flag is encoded by the TI flag encoding unit 107 and stored in the TI flag encoded data holding unit 108.

When outputting, the image code data is extracted from the image code data holding unit 104 and
At 05, the image data is decoded and passed to the output unit 110. On the other hand, the corresponding TI flag code data is extracted from the TI flag code data holding unit 108 and decoded by the TI flag decoding unit 109 to output the TI separation flag to the output unit 110.
Pass to The output unit 110 performs image processing for output on the image data with reference to the TI separation flag,
Output an image.

In such an apparatus, since the image code data and the TI flag code data are stored separately, it is necessary to manage the correspondence between the image code data and the TI flag code data. In order to simplify such management, it is conceivable to set the image data and the attribute data as one data string. When the image data and the attribute data are mixed, if the image data is encoded as it is, it can be encoded at one time, and the management of the encoded data becomes easy.

As one method of combining image data and attribute data such as a TI separation flag into one data string, a method of continuing attribute data after all image data is considered. However, according to this method, when performing the output processing in the output unit 110, the corresponding attribute data cannot be obtained until all the image data has been read. Therefore, this method is not suitable for the sequential processing and stores the image data. This is not a good idea because it requires a memory area.

As another method of combining the image data and the attribute data into one data string, there is a method of adding the attribute data to the original image data in pixel units or fixed block units. FIG. 26 is an explanatory diagram of an example of image data in which attribute data is mixed. FIG. 26A shows an example in which attribute data of the pixel is added before 1-bit pixel data. Here, the attribute data is 3
Is a bit. FIG. 26B shows an example in which 1-bit attribute data is added before data of 8 pixels. In such a data structure, attribute data can be associated with each pixel or every eight pixels.
The processing can be started in units of pixels or 8 pixels, and the memory area for storing these data can be reduced.

However, when data having such a data structure is encoded by the above-described method such as JBIG,
There is a problem that the code amount increases. For example, FIG.
In the case of the data structure shown in (A), if a template as shown in FIG. 15 is used, the reference pixels X ₀ and X ₁ when the pixel 0 is the target pixel are the second and third bits of the attribute data
This is a bit. However, the attribute data has almost no relation to the value of the pixel, and the value is set. Therefore, the attribute data having a weak correlation with the pixel 0 is referred to in the prediction of the pixel 0. For this reason, mispredictions increase, and the code amount increases. Further, even in a context where the efficiency of the reference pixels is all 0, etc., it is expected that many prediction errors occur, and there is a problem that the probability estimation state is shifted to a state worse than the steady state.

In order to efficiently compress images having different attributes in one image as shown in FIG. 26, for example, Japanese Patent Application Laid-Open No. Hei 6-38048 proposes a method of switching between a template and a prediction method by an attribute signal. Have been. However, this document has a problem that the entire code amount including the attribute signal becomes large because the attribute data is not compressed but simply inserted between the code data.

[0042]

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned circumstances, and includes a data string obtained by expanding a multi-valued image into a level plane, a data string in which image data and its attribute data are mixed, and the like. Provided is an encoding method and an encoding device capable of efficiently encoding a data string in which a plurality of types of data are mixed, and decoding code data encoded by such an encoding method and an encoding device. It is an object of the present invention to provide a decoding method and a decoding device.

[0043]

SUMMARY OF THE INVENTION An encoding method according to the present invention comprises:
For example, a data string in which two types of data of image attribute data and image data are mixed, or a data string in which a multi-valued image in which each pixel is composed of a plurality of levels of data is subjected to level plane development and each level is a different type, such as When encoding a data string in which types of data are mixed, a plurality of pieces of reference data of the same type that have already appeared according to the type of target data to be encoded are selected to generate a context. For example, when the data sequence is a data sequence in which image data and its attribute data are mixed, at least when the data of interest is image data, a plurality of reference data are selected from the already-existing image data. Since it is considered that there is a strong correlation between the image data, the selected reference data is considered to have a strong correlation with the target data.

Similarly, for example, when the data sequence is a data sequence in which a multi-valued image is developed into a level plane and each level is of a different type, data of the same type (ie, the same level) as the target data is selected as reference data. . In a multi-valued image, it can be statistically expected that the density change between a certain pixel of interest and surrounding pixels is small. For this reason, each pixel value may be considered to have a close value in many cases, and it is considered that adjacent pixels have almost the same value when viewed at each level. That is, when a certain level of a certain pixel is set as attention data, it is considered that data of the same level (kind) as the attention data of surrounding pixels and the attention data have a strong correlation.

As described above, in a data string in which a plurality of types of data coexist, attention is paid to the fact that there is a strong correlation between the same types of data. By selecting a plurality of types of reference data, data having a strong correlation with the target data can be selected as reference data, and a context can be generated. In general, it is known that when a value of attention data is predicted from reference data having a strong correlation, the hit rate is high. As described above, by selecting data having a strong correlation with the target data as reference data and generating a context, a prediction with a high hit rate can be performed. In the predictive coding, the code is configured so that the code amount is reduced when the prediction is correct. Therefore, the code amount can be reduced by improving the predictive accuracy, and the coding is efficiently performed. be able to.

Further, it is possible to prepare prediction data corresponding to each type, and switch between them according to the type of the data of interest. When predicting certain types of featured data,
The prediction data may be rewritten due to the mis-prediction, and the prediction of another type of target data may be incorrect. In such a case, misprediction causes an increase in the code amount. However, by switching the prediction data according to the type of the data of interest, the prediction can be performed without being influenced by other types of data. Therefore, the predictive accuracy can be improved for that type of data. , The code amount can be reduced as a whole.

It should be noted that, among attribute data in the case where the data string is a data string in which image data and its attribute data are mixed, considering that the attribute data is merely a code,
The correlation may be weak. Further, in the case where the data sequence is a data sequence in which a multi-valued image is developed on a level plane and each level is of a different type, a slight density change of the image between predetermined types of data including at least the level corresponding to the least significant bit. It is considered that the correlation is weak because the value changes depending on the value. For these types of weakly correlated data,
Even if prediction is performed, the hit rate is low, and conversely, there is a possibility that predicted data may be disturbed and the predictive hit rate may be reduced even for data having a high hit rate. Therefore, for data of a type having a weak correlation, a certain predetermined state is selected as reference data, and other prediction data is not disturbed apart from prediction data. As a result, data having a high hit rate can be predicted and coded without lowering the hit rate, so that the coding efficiency can be improved as a whole.

In the encoding apparatus of the present invention, when encoding a data string in which a plurality of types of data are mixed as described above,
In view of the fact that it is possible to improve the predictive accuracy and perform efficient encoding by selecting a plurality of reference data of the same type that have already appeared according to the type of the target data to be encoded, The generation unit selects a plurality of pieces of reference data of the same type that have already appeared according to the type of target data to be encoded, and generates a context. The prediction unit predicts the value of the data of interest from the context generated by the context generation unit, so that a prediction value with a high predictive accuracy can be output. In this way, the coding unit performs predictive coding using the predicted value having a high predictive accuracy. As described above, in the predictive coding, the code is configured so that the code amount is reduced when the prediction is correct, so that the code amount can be reduced by improving the predictive accuracy. An encoding device that generates code data with a small code amount can be obtained.

Also, by performing inverse prediction in the same manner as the above-described encoding method and encoding apparatus, it is possible to obtain a decoding method and a decoding apparatus capable of efficiently decoding encoded data. it can.

[0050]

FIG. 1 is a block diagram showing a first embodiment of the present invention. In the figure, the same parts as those in FIG. 14 are denoted by the same reference numerals, and description thereof will be omitted. 31 is input data, 32 is a context generation unit, 33 is an encoding unit,
41 and 42 are line memories, 43 is a shift register, 4
Reference numerals 4 and 45 denote a template extraction unit, 46 denotes a fixed context unit, 47 denotes a context selection unit, 48 denotes a control unit, and 49 denotes a control unit.
Is a control signal. In this example, the probability estimation table 6 and the arithmetic encoding unit 9 are collectively referred to as an encoding unit 31.
The context generation unit 32 changes the template according to the data to be encoded, generates a context 3 according to the template to be used, and passes the context 3 to the prediction unit 4.

FIG. 1B shows an example of the configuration of the context generating section 32. The line memories 41 and 42 hold data of a previous line and a line before the previous line in order to generate a template including a plurality of lines. The input data 31 is input to the line memory 41, and the data of the previous line is output. The line memory 42 is a line memory 41
, The data of one line before is output, and the data of two lines before is output with a delay of another line. The shift register 43 receives the input data 31, the data of the previous line output from the line memory 41, and the data of the previous two lines output from the line memory 42, and references the surroundings of the target data to be encoded. Holds data that can be data. Of course, the number of line memories may be increased or decreased to increase or decrease the number of lines to be referred to.

The template extracting units 44 and 45 extract an arbitrary part of the data stored in the shift register 43 and generate a prediction context. Different templates are set in the template extractor 44 and the template extractor 45, respectively, and a context is generated according to the set template. The fixed context section 46 always outputs a fixed predetermined context irrespective of the input data 31. The context output by the fixed context unit 46 is preferably a context having a low probability of being generated by the template extracting units 44 and 45. Thus, even if the prediction table in the prediction unit 4 is rewritten when the output of the fixed context unit 46 is selected, the template extraction unit 4
The prediction using the context output from 4, 45 can be minimized.

The context selecting section 47 receives a template signal from the template extracting section 44 according to a control signal 49 from the control section 48.
45, one of the outputs of the fixed context section 46 is selected and output as the context 3; Note that the selector may be provided before the template extracting units 44 and 45 and the fixed context unit 46. The control unit 48 outputs a control signal 49 to the context selection unit 47 according to the data format of the input data 31.

In this example, an example in which two template extraction units are provided is shown, but three or more template extraction units may be provided. Further, two or more fixed context parts 46 may be provided or may not be provided. Alternatively, a combination of one template extraction unit and a fixed context unit may be used.

An example of the operation according to the first embodiment of the present invention will be described. Input data 31 to be encoded
Is input to the line memory 41 and the shift register 43 of the context generation unit 32 and to the arithmetic coding unit 9. The line data 41 of the context generation unit 32 delays the input data 31 by one line, and the line memory 42 delays another line. Shift register 43
, Input data 31, input data one line before and input data two lines before are input. Shift register 43
Shifts the data in synchronization with the timing at which the input data 31 is sent, and holds three lines of data that can be reference data.

The template extracting sections 44 and 45 take out data from the data held in the shift register 43 in accordance with the template, and generate prediction contexts, respectively. In addition, fixed context part 4
6 outputs a predetermined context. The control unit 48 controls the control signal 49 according to the data format of the input data 31 so as to select a context composed of reference data having a strong correlation with the target data as much as possible according to the type of the target data to be encoded. To the context selector 47. Context selector 47
Selects one of the contexts generated by the template extraction units 44 and 45 or a predetermined context output from the fixed context unit 46 in accordance with the control signal 49 from the control unit 48, and
Is output to the prediction unit 4.

The operation after the prediction unit 4 is the same as the conventional operation, and the description is omitted here. However, as described above, the context 3 output from the context generation unit 32 is the same as the target data to be encoded. Since the reference data is composed of reference data having a strong correlation, the prediction value output from the prediction unit 4 has a high possibility of matching (predicting) with the input data 31 and the code amount of the code data 10 can be reduced. .

Next, the generation of a context in the context generation unit 32 will be further described based on some specific examples of the input data 31. First, input data 31
The case where data in which image data and attribute data are mixed is input will be described. FIG. 2 is an explanatory diagram of an example of a data format of input data in which attribute data is added to each pixel data and an example of a template to be used in the first embodiment of the present invention, and FIG. FIG. 5 is an explanatory diagram of data used for (1).
The data format of the input data 31 shown in FIG. 2A is the same as that shown in FIG. 26A. In this example, as the input data 31 shown in FIG. 2 (A), 1 to the previous bit of the pixel data d _mn, attribute data (a _mn of 3 bits indicating the attribute of the pixel, b _mn, c _mn ) Is added. The attribute data is, for example, a TI separation flag.

Input data 3 in which image data and attribute data in a data format as shown in FIG.
In order to efficiently predict 1, as shown in FIG. 2B, it is considered to be most efficient to use a template that refers to data every three bits. FIG.
In (B), P is the data of interest, and ○ indicates the position of the reference data. By generating the context by referring to the data at the position indicated by the circle in FIG. 2B, the context becomes a 10-bit context as in the context shown in FIG.

FIG. 3 shows a part of three lines of input data. Each data is given the same symbol as in FIG. 2A for identification. Now, in FIG. 3 (A), the as input data 31 is a target data to be encoded, and data d ₃₃ is inputted pixel data is the pixel data. At this time,
When the template shown in FIG. 2B is applied, FIG.
In (A), d ₁₂ , d ₁₃ , d ₁₄ , d ₂₁ ,
Each data _{d 22, d 23, d 24} , d 25, d 31, d 32 is referenced. These data are all pixel data and are considered to have a strong correlation. These data are the pixel data at positions X _{0 to} X ₉ in the template shown in FIG. 15 in the image data before the attribute data is inserted. Thus, when the input data is pixel data, the same context as before the attribute data is inserted can be created by using the template shown in FIG.

Next, in FIG.
It is assumed that the next data _a34 is input as 1. This data _a34 is the first bit of the 3-bit attribute data. In this case, when the template shown in FIG. 2 (B) is applied, a ₁₃ , a ₁₄ ,
Each data _{a 15, a 22, a 23} , a 24, a 25, a 26, a 32, a 33 is referred to. These data are all the first bits of the 3-bit attribute data. Similarly, when the second bit and the third bit of the attribute data are input as the input data 31, all the second bits of the attribute data can be obtained by applying the template shown in FIG. The third bit is referenced. When the attribute data is, for example, a TI separation flag, the bit string of the TI separation flag is often the same in a certain area, so that the data at the same bit position is likely to be the same. Therefore, the referenced data is considered to have a strong correlation. As described above, even when the input data is attribute data, by using the template shown in FIG. 2B, it is possible to create a context that is easily predicted.

As described above, when a data string having a data format as shown in FIG. 2A in which image data and attribute data are mixed is input, a template as shown in FIG. 2B is used. Thus, in both cases where the target data to be encoded is pixel data or attribute data, a context can be created by referring to data having a strong correlation with the target data. Therefore, the prediction accuracy of the prediction unit 4 is improved, and the code amount of the code data 10 can be reduced.

In this example, the same template shown in FIG. 2B can be used for both attribute data and pixel data. In such a case, the context selector 47
Only the template extracting unit for creating a context using the template shown in FIG. 2B is selected. FIG.
When only the input data of the data format shown in FIG. 2A is handled, one template extraction unit using the template shown in FIG. 2B is provided, and another template extraction unit, fixed context unit 46, context selection Part 4
7. The configuration may be such that the control unit 48 is not provided.

FIG. 4 is a diagram for explaining another example of a data format of input data in which attribute data is added to each pixel data and an example of a template to be used in the first embodiment of the present invention. FIG. 7 is an explanatory diagram of data used for generating a context. The data format of the input data 31 shown in FIG.
This is the same as that shown in FIG. In this example, as shown in FIG. 4A, 1-bit attribute data _ami is added to the head of each input data 31 for every eight pixels. The pixel data is indicated by _dmn .

Input data 3 in which image data and attribute data in a data format as shown in FIG.
The template for efficiently predicting 1 differs for each position of the pixel data. Basically, a predictable context can be obtained efficiently by using the template shown in FIG. 15, that is, the template shown in FIG. However, when this template is used for pixel data of 2 bits each on both sides of the attribute data, the attribute data is included as reference data. Since the attribute data has a weak correlation with the pixel data, if the attribute data is referred to including the attribute data to generate the context, the predictive accuracy may be reduced. Therefore, by switching and using templates as shown in FIGS. 4B to 4F according to the position of the pixel data, a context can be generated without referring to the attribute data. FIG.
In (B) to (F), P is the data of interest, and a circle indicates the position of the reference data.

FIG. 5 shows a part of three lines of the input data. Each data is provided with the same symbol as in FIG. 4A for identification. Now, in FIG. 5A, it is assumed that data d ₃₆ which is pixel data is input as input data 31 which is target data to be encoded. At this time,
Even if the basic template shown in FIG. 4D is applied, attribute data is not included in the template. Therefore, each pixel data is referred to using the basic template shown in FIG. 4D. Then, a context may be generated.

Next, pixel data d _{37 is used} as input data 31.
Is entered. Now a basic template, Figure 4
When the template shown in (D) is applied, attribute data a
₂₂ will be included. Therefore, as excluding this attribute data a _22, using the template shown in FIG. 4 (E).
As a result, the context can be generated by referring to only the pixel data without including the attribute data in the reference data.

Next, pixel data d _{38 is used} as input data 31.
Is input. If the basic template shown in FIG. 4D is applied, attribute data a ₁₂ and a ₂₂ are included, so the template shown in FIG. 4F is applied. Similarly, when the pixel data d ₃₉ is input, the template shown in FIG. 4B is applied, and when the pixel data d ₃₁₀ is input, the template shown in FIG. 4C is applied.

As described above, by changing the template to be applied according to the pixel position, when the target data is pixel data, the context is generated by referring only to the pixel data having a strong correlation without referring to the attribute data. can do. By performing prediction using the generated context, prediction can be performed with the same predictive accuracy as when attribute data is not mixed, and the overall code amount can be reduced.

When the input data 31 is attribute data,
Similarly to the template shown in FIG. 2B, by using a template that is referred to every eight data, a context can be generated by referring only to the attribute data. Alternatively, the output of the fixed context section 46 may be selected. For example, in the case of data in which one bit of attribute data is added to a larger number of pixels than in this example, for example, one or several lines, it is impossible to generate a template with only attribute data, and therefore, the fixed context part Preferably, a fixed context output by 46 is selected.

In this example, the pixel data shown in FIG.
The five templates shown in (B) to (F) are used. When a template is used for attribute data, another template is used. Therefore, five or six template extraction units may be provided for each template, and switched by the context selection unit 47 according to the input data 31. In addition,
When a template is used for attribute data, the fixed context section 46 may not be provided.

Next, a case where data obtained by expanding a multi-valued image into a level plane as the input data 31 will be described. FIG. 6 is an explanatory diagram of an example of a data format of input data expanded in a level plane and an example of a template to be used in the first embodiment of the present invention. FIG. FIG. Input data 3 shown in FIG.
The data format of 1 is basically the same as that shown in FIG. 24 described above, and this example shows an example in which 8 bits per pixel. Of the 8 bits, M
Let SB be level 7 and LSB be level 0, and call each bit position by level. It is assumed that the multi-valued image data is pure gradation data that has not been subjected to gray code conversion or the like.

Generally, in a multivalued image having many low frequency components such as a photographic image, it can be statistically expected that a change in density between a certain target pixel and surrounding pixels is small. Therefore, it can be considered that each pixel value takes a close value in many cases. When the change in pixel value is small and takes close values, in many cases only the value of the lower one or several bits near level 0 changes, and the upper bits including level 7 hardly change. Therefore, it is considered that the higher bits including the level 7 have a strong correlation with the surrounding pixels. That is, in the upper bits including level 7,
By referring to the same level of each pixel, a context in which prediction can be efficiently performed can be created.

Therefore, as shown in FIG. 6B, a template in which reference data is set every seven bits is applied. Thus, as shown in FIG. 7A, for example, when the data of interest is data of level 7, the reference data is all data of level 7. Similarly, when the data of interest is level 6 data, all the reference data is level 6 data, as shown in FIG. 7B. Similarly, for other levels, only data of the same level can be referred to. Thus, for the upper data, a context can be created by referring to data having a strong correlation, so that the predictive accuracy can be increased and the code amount can be reduced.

However, although the density change is small in the entire image, the value changes in the lower bits even with a small density change. As in this example, 256 gradations per pixel (=
In an 8-bit halftone image, a gradation change of less than 10 gradations actually occurs at random. As described above, it is not an exaggeration to say that the change in the value of the lower bits takes a random value almost in the same manner as noise in data of level 0, for example. That is, it is considered that the correlation between the lower-order bits and the data of the same level of the peripheral pixels is not so strong.

For example, a template as shown in FIG. 6B is applied to such lower-order bit data having a weak correlation, and efficient prediction can be performed by referring to the same lower-order data. Cannot be done. It is also difficult to create another efficient prediction template. In other words, no matter what template you create,
It is thought that it does not significantly affect the prediction efficiency. But,
As in the case of the data of the high-order bit including the level 7, when the template shown in FIG. 6B is applied, in some cases, the context frequently appears in the data of the high-order bit as the context for the low-order bit. The same context may be generated, in which case the prediction may be distorted and the predictive accuracy of the data of the upper bit may be reduced.

For example, assuming that the gray level is halftone data of about 60 to 90 continuously with light gray level halftone data, almost all levels 7 are 0 and all used contexts are 0. In this case, in the prediction at level 7, since the prediction continues, codes are hardly generated in the arithmetic coding method of JBIG. However, when the same template is used at level 0, it is quite possible that an all-zero context is accidentally generated, and furthermore, a prediction error is caused. In this case, STATE of the probability estimation table 6 for this context
Is changed in the direction of increasing the LSZ area, so that even if the prediction is successful in the context of all 0s thereafter, the code amount drops significantly compared to the case where the prediction continues.

Therefore, the context is fixed in the lower bits, and a template which is considered to be unlikely to be generated, for example, used in the JBIG TP process, is generated in the fixed context section 46 and used. . As a result, the prediction efficiency of the lower-order bit data itself does not increase so much, but does not adversely affect the prediction of other levels, so that the code amount of the entire image can be reduced. As the fixed context generated in the fixed context unit 46, for example, the value using each pixel position shown in FIG. 6C can be used.

It is optional to use the template shown in FIG. 6B at each level, or to select a fixed context as shown in FIG. 6C, for example.
For example, the template shown in FIG. 6B can be used for levels 7 to 1, and a fixed context can be used for level 0. in this case,
Creation of a context using the template shown in FIG. 6B is performed by the template extracting unit 44, and FIG.
The fixed context shown in FIG.
The context selection unit 47 is controlled so that the output of the template extraction unit 44 is selected when the level is 7 to 1 and the output of the fixed context unit 46 is selected when the level is 0 according to the level of the input data 31. I just need. In this case, the template extracting section 45 is unnecessary and need not be provided. Of course, the template extraction unit 45 may apply another template for the middle-order bit, and for example, for levels other than levels 0 and 7, the context selection unit 47 may select the output of the template extraction unit 45. Good.

FIG. 8 is a block diagram showing a second embodiment of the present invention. In the figure, the same parts as those in FIG. 34 is a prediction unit, 51,
52 is a prediction memory, 53 and 54 are selectors, and 55 is a control unit. In the second embodiment, an example is shown in which a plurality of prediction memories 51 and 52 are provided in the prediction unit 34, and a prediction memory to be used is selected according to the type of data. The context generation unit 32 includes line memories 41 and 42, a shift register 43, and a template extraction unit 44.
However, when generating a context in the template extracting unit 44, a template corresponding to the input data 31 is used.

The prediction memories 51 and 52 store prediction tables, for example, as shown in FIG. The prediction value and state corresponding to each context written in the prediction table are stored in the prediction memory 51.
And the prediction memory 52 may be different. Also, the number of prediction memories may be three or more.

The selector 53 selects either the output of the prediction memory 51 or the output of the prediction memory 52, and outputs a predicted value and a state. The selector 54 selects either the prediction memory 51 or the prediction memory 52 and outputs the new prediction value and state to be written back to the prediction table. The control unit 55 controls the selector 53 according to the type of the data of interest to be encoded, and selects either the prediction memory 51 or the output of the prediction memory 52. When the prediction table is changed by encoding, the selector 54 is controlled in accordance with the type of the coded data of interest to select a prediction memory in which the prediction table is changed.

An example of the operation according to the second embodiment of the present invention will be described. First, the case of input data in which pixel data and attribute data are mixed as shown in FIG. In the case of the input data of the data format shown in FIG. 2A, the context can be generated using the template shown in FIG. 2B for both the pixel data and the attribute data as described above. there were.
In this case, although the same template is used, it is expected that a context generated from the pixel data and a context generated from the attribute data will be different contexts due to a difference in data characteristics. However, for example, even when all of the attribute data is the same, the context generated from the pixel data becomes the same context as the context generated from the attribute data, and if the prediction is incorrect, the prediction of the next attribute data may also be incorrect. Therefore, the amount of code is remarkably increased due to such misprediction.

In order to avoid such a case, for example, when the target data is pixel data, the prediction memory 51 is used, and when the target data is attribute data, the prediction memory 52 is used.
The prediction result is switched by the selector 53 so that the prediction is performed using. Thus, even when the context generated from the pixel data and the context generated from the attribute data are the same, the prediction accuracy can be improved without affecting the mutual predicted values. As a result, the code amount can be reduced.

As another example, a case where data obtained by expanding a multi-valued image into a level plane as shown in FIG. 6A will be described. In this case, as described above, by using the template every seven bits shown in FIG. 6B, it is possible to generate a context by referring to the data for each level for all levels. Also, as described above, in the multi-valued image data developed in the level plane as in this example, the prediction is easily performed at the upper level, and the prediction is easily lost at the lower level, as described above. In the above-described first embodiment, an example is shown in which a fixed context is used at a lower level where prediction is likely to be lost. However, in the second embodiment, the same template is used, and the prediction accuracy of each level is used. The prediction memory is used separately according to. That is, for example, the prediction memory 51 is used for an upper level that is easy to predict, for example, levels 7 to 5, and the prediction memory 52 is used for a lower level that is difficult to be predicted, for example, levels 4 to 0. Then, the control unit 55 manages the level of the data of interest, and the selectors 53 and 54 select a prediction memory. As a result, the prediction table in the prediction memory 51 is not rewritten by the lower-level prediction that is difficult to predict, so that the upper level is easily influenced by the lower level without being affected by the lower level. It is. Therefore, the code amount can be reduced as a whole.

FIG. 9 is a block diagram showing a first modification of the second embodiment of the present invention. In the figure, the same parts as those in FIGS. 1 and 8 are denoted by the same reference numerals, and description thereof will be omitted. 56 is a prediction memory. In this modification, an example is shown in which a plurality of template extracting units 44 and 45 are provided in the context generating unit 32 as in the first embodiment described above, and a fixed context unit 46 is further provided. Of course, only a plurality of template extracting units may be provided, or three or more template extracting units may be provided. Further, in this example, the prediction memories 51, 52, and 56 corresponding to the template extracting units 44, 45 and the fixed context unit 46 are provided without providing the context selecting unit 47.

According to such a configuration, the context output from each of the template extracting units 44 and 45 and the fixed context unit 46 is input to the corresponding prediction memories 51, 52 and 56, and the predicted value and the state are stored. Is output. The selector 53 selects the output of the prediction memory according to the type of the input data 31 according to the control of the control unit 55. Thus, a context can be generated using a plurality of templates corresponding to the type of the input data 31, and a corresponding prediction memory can be used. Therefore, the prediction table of each prediction memory is used for prediction only for the context generated from the corresponding type of data without being affected by the context generated from other types of data at the time of prediction. The rate is increased, and the overall code amount can be reduced.

By using such a configuration, for example, as shown in FIG. 4A, attribute data is converted into input data of a data format in which image data and attribute data are mixed, and data obtained by expanding a multi-valued image into a level plane. Efficient encoding corresponding to a data string in which various different types of data such as mixed input data are mixed can be performed.

FIG. 10 is a block diagram showing a second modification of the second embodiment of the present invention. Reference numerals in the figure are the same as those in FIG. The second modification shows an example in which the fixed prediction value and the fixed state can be selected by the selector 53. For example, in the above-described first embodiment, the case where the fixed context unit 46 of the context generating unit 32 is selected is a case where the predictive accuracy is low. In such a case, the predictive predictive value does not improve so much regardless of the predicted value. For example, even if the prediction table is updated, the effect is hardly increased. Conversely, it is not necessary to update the prediction table, that is, even if a predetermined prediction value is always output, the prediction accuracy rate hardly changes. Therefore, in the second modification, instead of providing the fixed context section 46 in the context generation section 32,
The fixed prediction value and the fixed state are selected by the selector 53.

When the fixed prediction value and the fixed state are selected, rewriting data such as a new prediction value may be returned from the encoding unit 33. In such a case, it is not necessary to change the fixed prediction value and the fixed state. . In this case, the selector 54 may be controlled so that the rewrite data is not written to the prediction memory 51.

In the example shown in FIG. 10, only one prediction memory is shown, but a plurality of prediction memories may be used.
Further, similarly to the first modification, the context generation unit 32
It is also possible to provide a plurality of template extraction units in the server.
In this case, the fixed context section 46 is unnecessary.

The embodiments of the encoding method and the encoding apparatus according to the present invention have been described above. Even when the encoded data thus encoded is decoded, the same operation can be used to restore the original data. Can be.

FIG. 11 is a block diagram showing a third embodiment of the present invention. In the figure, the same parts as those in FIG. 20 are denoted by the same reference numerals, and description thereof is omitted. 61 is output data, 62 is a context generation unit, and 63 is a decoding unit.
In this example, the probability estimation table 6 and the arithmetic decoding unit 21 are collectively shown as a decoding unit 63. FIG. 11B shows a configuration example of the context generation unit 62, and this configuration is the same as the configuration of the context generation unit 32 shown in FIG. 1B. The only difference is that data input to the line memory 41 and the shift register 43 is output data 61.

The output data 61 decoded by the arithmetic decoding unit 21 is input to the line memory 41 and the shift register 43. The shift register 43 has a line memory 4
The output data delayed by one line and the line memory 41
The output data delayed by two lines at steps 42 and 42 is also input. The template extracting units 44 and 45 have different templates, respectively, refer to the data in the shift register 43 according to each template, and generate and output a context. The fixed context section 46 outputs a predetermined context. The context selection unit 47 selects one of the outputs of the template extraction units 44 and 45 and the fixed context unit 46 according to a control signal 49 from the control unit 48 corresponding to the type of the target data to be decoded, and Output to the prediction unit 22.

The template used by each of the template extracting units 44 and 45 is the same template used at the time of encoding. Further, the control unit 48 controls so that the template extraction units 44 and 45 or the fixed context unit 46 to which the same template as at the time of encoding is applied are selected according to the type of the data of interest.

For example, as shown in FIG. 2A, with respect to coded data obtained by coding data in a data format in which image data and attribute data are mixed, as shown in FIG. Using the template shown, a context is created from the decoded output data and prediction is performed. Further, if it is code data obtained by encoding data in the data format as shown in FIG. 4A, five or six template extracting units are provided, each of which is provided with a template extracting part as shown in FIGS. The context may be generated by using any of the attribute data templates as needed. Further, in the case of encoded data obtained by encoding data obtained by level-expanding a multi-valued image as shown in FIG. 6, for example, the template extracting unit 44 creates a context using the template shown in FIG. What is necessary is just to comprise so that the context shown in FIG.6 (C) may be output from the fixed context part 46. FIG. In the case of other data formats as well, it may be configured so that the use of a template or selection of a fixed context corresponding to the encoding is performed in the same manner.

The context output from the context generation unit 32 is input to the inverse prediction unit 22, the inverse prediction value is output to the arithmetic decoding unit 21, and the state is output to the probability estimation table 6.
Output to The probability estimation table 6 reads information such as LSZ and SW on the basis of the state and outputs the information to the arithmetic decoding unit 21.
The arithmetic decoding unit 21 decodes the code data 10 from the information and the inverse prediction value, and outputs the decoded data as output data 61. Further, if necessary, rewrite data 24 such as a new inverse prediction value and the rewrite signal 12 are sent to the inverse prediction unit 22 to rewrite the inverse prediction table.

By performing prediction and decoding in the same manner as at the time of encoding, data in which image data and attribute data are mixed, data in which a multi-valued image is expanded into a level plane, and the data and attribute data are mixed. The original data can be restored by decoding the encoded data obtained by encoding the decoded data.

At the time of decoding, the same configuration as that of the above-described second embodiment and the first and second modified examples can be adopted. That is, the inverse prediction unit 22 in FIG.
Is configured in the same manner as the prediction unit 34 shown in FIG. 8 or FIG.
The context generation unit 32 and the inverse prediction unit 22 can be configured similarly to the context generation unit 32 and the prediction unit 34 shown in FIG.

FIG. 12 is a block diagram showing an application of the present invention. In the figure, the same parts as those in FIG. 111 is a mixing unit, 112 is an encoding unit, 113 is a code data storage unit, 114 is a decoding unit, and 115 is a separation unit. The image data input from the input unit 101 is processed by the image processing unit 102 and input to the mixing unit 111, and the TI data
The separation processing is performed, and the TI separation flag is input to the mixing unit 111. In the mixing unit 111, for example, FIG.
As shown in (A), data in which image data and a TI separation flag are mixed is created and passed to the encoding unit 112. The encoding unit 112 encodes the data in which the image data and the TI separation flag are mixed by the configuration and the method described in the above-described first or second embodiment of the present invention.
13 is stored. At this time, the encoding unit 112 uses a different template for the image data and the TI separation flag as described above, selects a fixed context for the TI separation flag, and uses a different prediction table for each. By performing the prediction processing in this way, the code amount is reduced.

At the time of output, the code data is extracted from the code data storage unit 113, decoded by the decoding unit 114, and returned to the data in which the image data and the TI separation flag are mixed, and the separation unit 15 separates the image data and the TI separation flag. And passes it to the output unit 110. The output unit 110 forms an image by performing image processing for output on the image data while referring to the TI separation flag.

In the configuration shown in FIG. 12, image data and T
Since the I separation flag is mixed and encoded, the management of the code data stored in the code data storage unit 113 is easy. In addition, since the data into which the TI separation flag is inserted is encoded for each data of one or several pixels, the output unit 110 can perform processing while decoding the data, and the buffer capacity of the output unit 110 can be reduced. High-speed output processing is possible. In the case of encoding data in which such image data and the TI separation flag are mixed, the amount of code has been extremely large in the related art. However, by using the present invention, the code is efficiently encoded and the amount of code is reduced. be able to. Therefore, the capacity of the code data storage unit 113 per image can be reduced, and the code data storage unit 1
13 or a large amount of images can be stored.

[0103]

As apparent from the above description, according to the present invention, attribute data and image data are mixed,
Or, data obtained by expanding a multi-valued image to a level plane, etc.
When encoding a data string in which a plurality of types of data are mixed, a context is generated by selecting a plurality of reference data of the same type that has already appeared according to the type of the target data, and a context of the target data is generated from the generated context. A value is predicted, and prediction encoding is performed based on the predicted value and the value of the data of interest. The surrounding data of the same type is considered to have a strong correlation, and by making use of the fact that a prediction is performed using data with a strong correlation, the hit rate is good. Rate can be improved. Normally, since predictive coding is configured to reduce the amount of code data when a prediction is correct, the code amount is reduced by improving the predictive accuracy as described above, and efficient coding is performed. Can be performed.

Further, for a type of data having a low prediction efficiency, a predetermined state is selected as reference data, or another prediction table is used, so that the prediction efficiency of the data having a low prediction efficiency is reduced. By encoding good data so as not to be affected, and encoding the types of data with good prediction efficiency so that the code amount does not increase, the total code amount can be reduced.

As described above, since a data sequence in which a plurality of types of data are mixed can be efficiently encoded, it is not necessary to separately encode each type, and it is possible to efficiently encode as one encoded data. There is an effect that can be.

Further, as an encoding apparatus for realizing the above-described encoding method, the context generation unit generates a context by selecting a plurality of reference data of the same type that have already appeared according to the type of the data of interest. , It is possible to output a context composed of the same type of data strongly correlated with the target data, and it is possible to output a predicted value with a high predictive accuracy by using this context in the prediction unit. The encoding unit performs predictive encoding based on the predicted value and the data of interest. However, since the predictive accuracy of the predicted value is high, it is possible to generate code data with a small code amount. Also, a plurality of prediction memories or a fixed context section is provided,
For a type of data having a poor predictive accuracy, a predetermined state is selected as reference data in the fixed context part,
It can be configured to use a different prediction memory from other types of data. By this, the type of data having a good predictive accuracy is not affected by the prediction of the type of data having a bad predictive accuracy, and the code amount is not increased for the type of data having a good predictive accuracy. By encoding, the entire code amount can be reduced.

Also, by performing inverse prediction in the same manner as in the above-described encoding method and encoding apparatus, it is possible to obtain a decoding method and a decoding apparatus capable of efficiently decoding encoded data. it can.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is an explanatory diagram of an example of a data format of input data in which attribute data is added to each pixel data and an example of a template to be used in the first embodiment of the present invention.

FIG. 3 is an explanatory diagram of data used for generating a context in an example of input data in which attribute data is added to each pixel data in the first embodiment of the present invention.

FIG. 4 is an explanatory diagram of another example of a data format of input data in which attribute data is added to each pixel data and an example of a template to be used in the first embodiment of the present invention.

FIG. 5 is an explanatory diagram of another example of data used for generating a context in another example of input data in which attribute data is added to each pixel data in the first embodiment of the present invention.

FIG. 6 is an explanatory diagram of an example of a data format of input data expanded in a level plane and an example of a template to be used in the first embodiment of the present invention.

FIG. 7 is an explanatory diagram of data used for generating a context in an example of input data that has been subjected to level plane development in the first embodiment of the present invention.

FIG. 8 is a block diagram showing a second embodiment of the present invention.

FIG. 9 is a block diagram showing a first modification of the second embodiment of the present invention.

FIG. 10 is a block diagram showing a second modification of the second embodiment of the present invention.

FIG. 11 is a block diagram showing a third embodiment of the present invention.

FIG. 12 is a block diagram showing an application example of the present invention.

FIG. 13 is a conceptual diagram of arithmetic coding.

FIG. 14 is a schematic diagram of a predictive coding method adopted in JBIG.

FIG. 15 is an explanatory diagram of an example of a template used in JBIG.

FIG. 16 is an explanatory diagram of an example of a prediction table stored in the prediction unit 4.

FIG. 17 is an explanatory diagram of an example of the probability estimation table 6.

FIG. 18 is an explanatory diagram of a specific example of input image data.

FIG. 19 is an explanatory diagram of a specific example of a prediction section table after rewriting.

FIG. 20 is a schematic diagram of a predictive decoding method adopted in JBIG.

FIG. 21 is an explanatory diagram of a basic operation of the arithmetic decoding unit.

FIG. 22 is an explanatory diagram of bit plane development.

FIG. 23 is an explanatory diagram of level plane development.

FIG. 24 is an explanatory diagram of an example of a data string that has undergone level plane development.

FIG. 25 is a block diagram showing an example of a case where intermediate information is stored with attribute data in a conventional image input / output device.

FIG. 26 is an explanatory diagram of an example of image data in which attribute data is mixed.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Image data, 2 ... Context generation part, 3 ... Context, 4 ... Prediction part, 5 ... State, 6 ... Probability estimation table, 7 ... LSZ and SW, 8 ... Prediction value, 9 ... Arithmetic coding part, 10 ... Code data, 11 ... rewrite data, 1
2 rewrite signal, 21 arithmetic decoding unit, 22 inverse prediction unit, 23 inverse prediction value, 24 rewrite data, 31 input data, 32 context generation unit, 33 encoding unit, 41, 42 ... Line memory, 43 shift register, 44, 45 template extraction section, 46 fixed context section, 47 context selection section, 48 control section, 49 control signal, 34 prediction section, 51, 52 prediction memory, 53, 54 ... selector, 55 ... control unit, 56 ...
Prediction memory, 61 output data, 62 context generation unit, 63 decoding unit, 101 input unit, 102 image processing unit, 103 image coding unit, 104 image code data holding unit, 105 image decoding Section, 106 ... TI separation section,
107: TI flag encoding unit, 108: TI flag code data holding unit, 109: TI flag decoding unit, 110 ...
Output unit, 111: mixing unit, 112: encoding unit, 113 ...
Code data storage unit, 114: decoding unit, 115: separation unit.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) H03M 7/36 H04N 1/417 H04N 7/32

Claims (11)

(57) [Claims] In an encoding method for encoding a data string in which a plurality of types of data are mixed, a plurality of reference data of the same type that has appeared according to the type of target data to be encoded is selected. A coding method comprising: generating a context; predicting a value of data of interest from the generated context; and performing predictive coding based on the predicted value and the value of the data of interest. 2. The encoding method according to claim 1, wherein when predicting the value of the data of interest, the prediction is performed using data for prediction corresponding to the type of the data of interest. 3. The encoding method according to claim 1, wherein the data sequence is a data sequence in which two types of data of image attribute data and image data are mixed. 4. When the data of interest is attribute data, a predetermined state is selected as reference data, and when the data of interest is image data, a plurality of reference data are selected from already appearing image data. 4. The method according to claim 3, wherein
Encoding method. 5. The data sequence according to claim 1, wherein the data sequence is a data sequence in which a multi-level image in which each pixel is composed of a plurality of levels of data is developed into a level plane and each level is of a different type. Coding method as described. 6. When the type of the data of interest is a predetermined type including at least a level corresponding to a least significant bit, a predetermined state is selected as reference data, and the type of the data of interest is other than the predetermined type. 6. The encoding method according to claim 5, wherein in the case of, the same type of data as the target data in a plurality of pixels located around the target pixel including the target data is selected as reference data. 7. An encoding apparatus for encoding a data string in which a plurality of types of data are mixed, wherein a context for selecting a plurality of reference data of the same type that has already appeared according to the type of target data to be encoded. A generation unit; a prediction unit that predicts a value of the data of interest from the state of the reference data selected by the context generation unit and outputs a predicted value; And a coding unit for performing predictive coding. 8. The apparatus according to claim 1, wherein the prediction unit includes a plurality of prediction memories each storing different data for prediction, and switching means for switching the prediction memory according to a type of the data of interest. Item 8. The encoding device according to Item 7. 9. The encoding apparatus according to claim 7, wherein the context generation section includes a fixed context section that outputs a predetermined state as reference data when the target data is of a specific type. apparatus. 10. A decoding method for decoding code data obtained by coding a data string in which a plurality of types of data are mixed, wherein a plurality of reference codes of the same type already decoded according to the type of target data to be predicted and decoded next. A decoding method comprising selecting data, generating a context, inversely predicting the value of the data of interest from the generated context, and performing predictive decoding based on the inversely predicted value and the value of the data of interest. 11. A decoding device for decoding code data obtained by coding a data sequence in which a plurality of types of data are mixed, a decoding unit for predictively decoding code data from the code data and its prediction value, A context generating unit that selects a plurality of reference data of the same type that has been decoded according to the type of the data of interest to be predicted and decoded; and A decoding device comprising: a prediction unit that predicts and outputs a prediction value to the decoding unit.