JP3075049B2 - Image coding device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION This invention relates to image picture data encoding device.

[0002]

2. Description of the Related Art Markov model coding is effective as a binary image coding method in facsimile and the like, and a technique utilizing this is disclosed in Japanese Patent Application Laid-Open No. 2-305225. Hereinafter, conventional Markov model coding will be described.

Here, if P (x _i ) is the probability of occurrence of an information source symbol x _i in a certain information source , the entropy (average information amount) of the information source is defined by equation (1).

[0004]

(Equation 1)

[0005] This equation means that taking the average weighted by the probability of occurrence of each symbol of the information amount -log ₂ P (x _i) obtained by learning that symbol x _i occurs. In the case of a binary image (x _i = 0 or 1), entropy is represented by equation (2).

[0006]

(Equation 2)

Further, it is generally known that this value indicates a theoretical compression limit that cannot be encoded with a code amount smaller than this when a code is assigned while focusing on only one pixel to be encoded. Therefore, in encoding, it is necessary to perform entropy encoding suitable for the occurrence probability P (x _i ) of each symbol so that the code amount approaches the entropy H as much as possible.

In general, the probability of a pixel having a value of 0 or 1 in image data depends on what the values of a plurality of pixels preceding it are. When the information source depends on the values of the preceding n pixels (hereinafter referred to as reference pixels), the information source is referred to as an n-fold Markov information source. The entropy of the n-fold Markov information source, that is, the entropy (conditional entropy) after knowing the n pixel values (x _i−1 ,..., x _in ) preceding the pixel of interest (x _i ) is expressed by the formula ( It is represented by 3). Here, P (x _i−1 ,..., X _in ) indicates that the value of the reference pixel is x
_i−1 ,..., x _in (joining probability), P (x _i /
x _i−1 ,..., x _in ) indicates that the value of the reference pixel is x _i−1 ,.
_{When “in”} , the probability (conditional probability) that the pixel value of interest becomes _xi is shown.

[0009]

(Equation 3)

In the relationship between the conditional entropy H ′ and H in the equation (2), H ′ ≦ H (4) holds, and the equality holds when the information source is not a Markov information source.

[0011] Here, the n-number of the 2 ⁿ possible states of the reference pixels to be referred to as the state of the Markov information source, representing each state S _i. Then, when formula (3) is transformed, formula (5) is obtained.
Is obtained.

[0012]

(Equation 4)

This is because the entropy h (s _i ) in each state
= -ShigumaP a _{(x i / s i) log} 2 P (x i / s i) in which it averaged and weighted by occurrence probabilities P (s _i) of each state. During encoding, for each state, the probability the target pixel takes 0 or _{1 P (x i / s i} ) the amount of codes in each state by performing the entropy coding suitable for the entropy h (s _i ), And the code amount as a whole can also approach H ′ in equation (5). Therefore,
The reference pixels need to be selected such that the conditional entropy is reduced. In this manner, a coding method that classifies the target pixel according to the state of the reference pixel and performs appropriate entropy coding for each state is called Markov model coding.

An example of Markov model coding is disclosed in
Arithmetic code type MELCO described in −305225
The DE will be described. FIG. 2 shows the configuration diagram. 11
Denotes reference image data to be coded, 12 denotes a reference pixel pattern creating means for selecting a reference pixel to be used for Markov state classification from input pixels before the target pixel and creating a reference pixel pattern, and 18 denotes a reference pixel pattern. Predicted value / predicted coincidence determining means for calculating the predicted value of the target pixel and the hit rate (prediction coincidence) of the prediction using the reference pixel pattern 13; 19 stores the predicted value and the predicted coincidence for the reference pixel pattern Predicted value / predicted match rate reference table, 20 is a predicted value (0 or 1), 21 is a predicted match rate, 22 is an encoding target symbol indicating match / mismatch between the predicted value and the pixel of interest, 2
Numeral 3 denotes an arithmetic encoder for encoding a symbol to be encoded according to the prediction coincidence rate 21, and numeral 24 denotes a code output from the arithmetic encoder.

[0015]

[Table 1]

Next, the operation of this encoder is shown in Table 1 below.
A description will be given using a value image as an example. In this example, 31 in FIG. 3 is the pixel of interest x _i , and 32 is the reference pixels x _i-1 and x _i-2 .
Since the reference pixel is 2 pixels as condition of Table 1 at four s ₀ ~s _3, probability P (s _i) of each state, the probability P that the pixel of interest becomes zero in each state (0 / s _i) , The probability P of 1
(1 / s _i ) is the value in Table 1. In this case, since the probability P (0) = 0.47 that the target pixel is 0 and the probability P (1) = 0.53 that the target pixel is 1, H = 0.
997 bits are obtained. As described above, this value is a theoretical compression limit when a code is assigned while paying attention to only one pixel to be encoded. In addition, the conditional entropy calculated from Equation (5) by classifying the target pixel based on the state s _i of two pixels x _i−1 and x _i−2 is H ′ = 0.72 bits. As described above, by classifying into Markov states and performing appropriate encoding for each state, H ′ = 0.72 bits
It is possible to perform encoding with a code amount close to.

At the time of encoding, first, the reference pixel pattern creating means 12 selects the pixels x _i-1 and x _i-2 from the encoded pixels to create the reference pixel pattern 13. Based on the reference pixel pattern 13, the predicted value 20 and the predicted matching rate 21 are read from the predicted value / predicted matching rate reference table.
If the state of the reference pixels if was s _1, the predicted value from Table 1 = 1, the predicted match rate = 0.6. Then, the prediction matching rate 21 is input to the arithmetic encoder 23.

If the value of the pixel to be coded is 1, the prediction was successful and the symbol 0 indicating that the prediction was successful is input to the arithmetic encoder 23 as the symbol 22 to be coded. I do. If the pixel value to be encoded is 0, the prediction is missed, and the symbol 1 indicating that the prediction was missed is input to the arithmetic encoder 23 as the encoding symbol 22. In the arithmetic encoder 23, the encoding target symbol 2 is encoded using an encoding parameter suitable for the prediction coincidence rate.
By encoding 2, encoding can be performed with substantially the same code amount as entropy.

[0019] At this time, if at k (bit) requires each state to store the prediction value and prediction matching rate, the magnitude of the predicted value and prediction match rate reference table 19, the number of pixels reference pixels Since there are two pixels, it is k × 2 ² bits.

[0020]

In Markov model coding, entropy generally decreases when the number of reference pixels is increased and the number of states is increased, but a reference table 19 storing predicted values and prediction coincidence rates for each state. Has a problem that it is difficult to greatly increase the number of reference pixels because the size of the reference pixel increases exponentially with the number of reference pixels.

In Table 1, the target pixel is classified into four states according to possible values of the reference pixel. However, it is considered that some of the four states are collectively set as one state. Hereinafter, this processing is referred to as Markov state degeneration, or simply degeneration. For example, if the states s ₁ and s ₂ in Table 1 are combined into one state, the states are reduced to _three states s ′ _{0 to} s ′ ₃ shown in Table 2.

[0022]

[Table 2]

When the states are degenerated, the entropy generally increases. However, when the states having close predicted coincidence rates are degenerated, the entropy is slightly suppressed. In the degeneration from Table 1 to Table 2, degeneration is achieved without increasing entropy because the predicted matching rates of the two degenerated states are equal. An object of the present invention is to realize a Markov model coding apparatus that performs such efficient degeneration.

[0024]

The Markov state reduction means according to the present invention predicts a value indicating the property of a pixel of interest,
By bringing together near the predicted level state into one state, and performs degeneracy of Markov states.

In the encoder of the present invention, the prediction target is the value of the target pixel or the black pixel appearance probability of the target pixel, and the target pixel is classified (that is, the state is degenerated) based on the value.

In the encoder according to the present invention, a plurality of prediction means and prediction level classification means are prepared, and the Markov state is degenerated by adaptively selecting them in the screen according to the properties of the reference pixels. It is.

[0027]

According to the first to third aspects of the present invention, a value reflecting the property of a reference pixel is obtained by performing prediction on a target pixel value, a black pixel appearance probability, and the like. As a result, the Markov state is degenerated with the increase in entropy suppressed. According to the fourth aspect of the present invention, the prediction means and the prediction level classification means in the prediction for the target pixel value, the black pixel appearance probability, etc. according to the first to third aspects of the present invention are arranged according to the state of the reference pixel. By performing the switching, the Markov state is degenerated while suppressing an increase in entropy.

[0028]

[Embodiment 1] Hereinafter, an embodiment of the present invention will be described. FIG. 1 shows an arithmetic code type MELCODE using the present invention.
FIG. In FIG. 1, reference numeral 14 denotes a reference pixel pattern 13 created by the reference pixel pattern creation means 12.
Pixel prediction means for predicting the pixel value of interest using
Is a prediction level classifying unit that quantizes the prediction level 15 calculated by the prediction by the target pixel prediction unit 14 and classifies states close to the prediction level into one state collectively. Reference numeral 18 denotes a classification result 17 obtained by the prediction level classification unit 16. Is a prediction value / prediction match rate determination unit that reads out a prediction value and a prediction match rate from the prediction value / prediction match rate reference table 19 in accordance with (1). Only the above-described parts are different from those in FIG. 2, and the description of the other parts will be omitted.

Next, the operation will be described using a binary image having a probability distribution shown in Table 1. First, using the reference pixels x _i−1 and x _i−2 of the reference pixel pattern 13, the target pixel prediction unit 14 calculates a prediction level y _i 15 from Expression (6). This y _i takes a real value and is called a prediction level to distinguish it from the predicted value 20 (0 or 1). y _i = 0.52x _i-1 + 0.44x _i-2 (6)

This prediction function is designed in advance so that the mean square error between the pixel value of interest and the prediction level is minimized as shown below. When e _i is a prediction error and a _i is a prediction coefficient, the prediction error e _i is represented by the following equation. In _{e i = y i -x i =} a 1 · x i-1 + a 2 · x i-2 -x i (7) wherein, if the E [e _i ^2] and mean squared prediction error e _i , E [e _i ² ] can be obtained by calculating the prediction coefficients a ₁ and a ₂ that satisfy Expression (8). ^{_{∂E [e i 2] / ∂a}} 1 = 0 ∂E [e i 2] / ∂a 2 = 0 (8) Equation (8) can be written as the following equation. This is Yule-Wa
It is called the lker equation.

[0031]

(Equation 5)

The prediction coefficient of the equation (6) is obtained by solving the equation (9) for the binary image of the probability distribution of Table 1.

[0033]

[Table 3]

[0034]

[Table 4]

[0035]

[Table 5]

The prediction level 15 thus obtained
Is y _i in Table 1. Next, as shown in Table 3, the prediction level y _i
By quantizing to a level, states having close prediction levels are combined into one, and the states are degenerated. Due to this degeneration, the conditional entropy becomes 0.78 bits. This value is higher than that of the entropy H = 0.72 bit shown in Table 1 due to the degeneration. However, Table 4
Alternatively, as shown in Table 5, the reference pixel is x _i-1 or x _i-2
Is smaller than the entropy when only one pixel has the same number of states, and the effect of the degeneration of the present invention can be confirmed. That is, the predicted value 20 corresponding to each state shown in Table 3,
The predicted matching rate 21 is converted to a predicted value / predicted matching rate reference table 19.
, And by performing arithmetic coding, it is possible to perform coding with a smaller code amount than in the case of Table 4 or Table 5 having a prediction value / prediction matching rate reference table 19 of the same size.

Embodiment 2 FIG. In the first embodiment, a linear prediction function is used as the pixel-of-interest prediction means 14. However, for example, a hierarchical neural network shown in FIG. 4 may be used as the prediction function. Each layer is called an input layer, an intermediate layer, and an output layer in order from the left of the figure, and each layer is composed of several units 43. Each unit in the input layer sends the input signal 41 to the next layer as it is, and in the intermediate layer and the output layer, each unit calculates the output value of each unit according to its value by taking the weighted sum of the output of each unit in the previous layer. I do. Then, the output of the unit in the output layer becomes the output 42 of the neural network. What determines the input / output characteristics of the neural network is a weighting factor when the weighted sum of outputs of the previous layer is obtained, and the weighting factor has a different value for each connection. The weight coefficient is determined in advance by a process called learning.

Next, the operation will be described. Since the operation is the same as that of the first embodiment except that a neural network is used as the prediction function, the description of the other operations is omitted. First, the reference pixel pattern 13 is input to the neural network of FIG. 4, and the prediction level 15 of the pixel value of interest is output from the neural network. Then, the prediction level 15 is classified by the prediction level classification means 16 to determine the state.

The weighting factor of the neural network at this time is
It is determined in advance by a back propagation learning rule so as to minimize the mean square error between the target pixel and the output value.
This is an operation of repeatedly inputting an input signal and correcting the weight coefficient so that the output approaches a desired value as an output called a teacher signal. For example,
In the case of the binary image having the probability distribution shown in Table 1, the reference pixel value (x _i-1 , x _i-2 ) in each state and the teacher signal x _{i corresponding} thereto
Are sequentially given to the neural network, and learning is repeated until the square error between the output value and the teacher signal converges. That is, the teacher signals in each state S ₀ , S ₁ , S ₂ , S ₃ are 0,
1,1,1.

[0040]

[Table 6]

The output y _i of each state of the neural network constituted by the weight coefficients determined in this way is shown in Table 6. If this is input to the prediction level classification means 16 shown in Table 3, it is possible to degenerate the Markov state as in the first embodiment.

Embodiment 3 FIG. In the first and second embodiments, the target pixel value is set as the prediction target. However, the probability that the target pixel becomes a black pixel (black pixel appearance probability) may be predicted. For example, when the learning of Example 2 in neural networks intended for binary images in Table 1, was used target pixel value x _i is a prediction target as a teacher signal, a black pixel appearance probability P (1 instead / S _i )
Is used. That is, the teacher signals in the states S ₀ , S ₁ , S ₂ , and S ₃ are 0.2, 0.6, 0.6, and 0.9, respectively.
Becomes

Embodiment 4 FIG. In the first to third embodiments, only one type of quantizer is used to classify the prediction function and the prediction level. However, a plurality of quantizers are prepared, and adaptively in the screen according to the reference pixel. You may switch.

[0044]

[Table 7]

[0045]

[Table 8]

The reference pixels are the three pixels x _i-1 and x
_i-2 and xi _-3 . The binary image at this time is shown in Table 7
It is assumed that the probability distribution shown in FIG. The conditional entropy H ′ when classified into the states S _{0 to} S _{7 based} on the value that the reference pixel can take is H ′ = 0.77 bits according to the equation (4). In the fourth embodiment, first, the prediction level 1 is calculated by the equation (10).
5 is calculated.

[0047]

(Equation 6)

In the case of the image shown in Table 7, the prediction level 15 of each state is y _{i in} Table 7. Next, as shown in Table 8, the prediction level 15 is quantized by different quantizers according to the value of x _i−3 . That is, when x _i−3 = 0, the threshold value is 0.4
In the case of x _i−3 = 1, the data is classified into two levels with a threshold value of 0.5. The difference from the first to third embodiments is that the reference pixel x
That is, the prediction function and the quantizer are switched according to the value of _i-3 . Thereby, the binary image of Table 7 is classified into the four states of Table 8, and the conditional entropy H ′ = 0.80 bit
Becomes

[0049]

[Table 9]

Table 9 shows the probability distribution and the conditional entropy when classified into four states according to the possible values of x _i-1 and x _i-3 . Despite being classified into the same four states as in Table 8, the entropy is higher than in Table 8, and the effect of Example 4 can be confirmed.

[0051]

According to the present invention, the conditional entropy of the image data can be reduced without greatly increasing the size of the encoding parameter reference table, and the code amount can be reduced.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an arithmetic code type MELCODE using the present invention.

FIG. 2 is a configuration diagram of a conventional arithmetic code type MELCODE.

FIG. 3 is a diagram illustrating a pixel arrangement of a target pixel and a reference pixel in a conventional example and Examples 1 to 3 of the present invention.

FIG. 4 is a diagram showing a structure of a neural network according to Embodiments 2 and 3 of the present invention.

FIG. 5 is a diagram illustrating a pixel arrangement of a target pixel and a reference pixel according to a fourth embodiment of the present invention.

[Explanation of symbols]

 14 Target pixel prediction means 15 Prediction level 16 Prediction level classification means 18 Prediction value / prediction match rate determination means 19 Prediction value / prediction match rate reference table 31 Target pixel 32 Reference pixel

Claims (4)

(57) [Claims] 1. A predicting means for predicting a value indicating the property of a target pixel using a coded reference pixel around the target pixel, and a target pixel having a similar prediction level output from the predicting means. One of the provided prediction level classification means for classifying a Markov state, the estimated hand relative to the pixel of interest
The prediction level determined by the prediction level output from the stage.
An image coding apparatus based on Markov model coding, wherein entropy coding suitable for each classification is performed according to the prediction level classification result of the classification means . 2. A picture coding device of claim 1 wherein, characterized in that the target pixel value prediction target. Wherein the probability of the target pixel prediction target becomes black pixels (hereinafter, a black pixel appearance probability) picture image encoding apparatus of claim 1 wherein, characterized in that a. 4. The apparatus according to claim 1, further comprising a plurality of said prediction means and said plurality of prediction level classification means, wherein said prediction means and said prediction level classification means are adaptively selected in a screen according to a state of said reference pixel. Claims 1, 2 and 3
Picture coding device according to claim.