JPS59185487A - Adaptive forecasting encoding and decoding system and device of multilevel picture signal - Google Patents

Abstract

PURPOSE:To raise an estimating accuracy, and to elevate forecasting encoding efficiency by considering the spatial distribution state of an optimum forecasting function corresponding to an encoded picture element, and estimating the optimum forecasting function in a present picture element time. CONSTITUTION:In a present picture element time, the information for showing an optimum forecasting function of a preceded picture element A and preceded line picture elements B, C is supplied in 3-bit parallel to a deciding circuit 21 by a signal line 52. This deciding circuit 21 determines an in-frame forecasting or an inter-frame forecasting the be used for a present picture element, basing on an estimating method of the optimum forecasting function to the present picture element, from a signal for showing a generating state of the optimum forecasting function in a reference picture element being an output of a storing circuit 19, and supplies a control signal to a selector 14. On the other hand, a forecasting error signal quantized by a quantizer 15 is inputted to a code converter 16, to encode to variable length and the supply to a transmission line and a recording medium.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a predictive coding/decoding system and device for multilevel image signals, and particularly to a predictive coding/decoding system and device that adaptively switches a plurality of prediction functions according to the characteristics of a screen. .

In the prediction of multivalued image signals such as television signals, (1) in addition to the fixed use of one prediction function, (2) the fixed combination of multiple prediction functions, ( 3) Various methods have been considered, such as adaptively switching between multiple predictions depending on the characteristics of the screen.

In this third adaptive prediction method, which performs prediction by adaptively switching prediction functions, a block is configured in units of multiple pixels, and the optimal prediction function is detected for each block, and the optimal prediction function used is This method transmits the code indicating the current value and the prediction error at the same time, and estimates the optimal prediction function for the next pixel by examining the local properties of the image signal using only information corresponding to already encoded pixels. There is a method.

In the latter case, similar estimation is possible on the receiving side, so there is no need to transmit a code indicating the used prediction function, and the circuit configuration is that much simpler. This invention belongs to this latter method.

If only the determination result at the current pixel time is used to estimate the optimal prediction function at the next pixel time, prediction errors that are different from those of surrounding pixels, for example, even if the surrounding prediction errors are small, Due to so-called isolated points where only one point has a large error regardless of the
The estimation of the optimal prediction function may be incorrect. This phenomenon tends to occur with signals on which random noise is superimposed. Even in such cases, if the generation state of the optimal prediction function used is understood two-dimensionally, accurate estimation is possible without being influenced by the existence of isolated points.

This point will be explained with reference to FIG. Here, it is assumed that two types of prediction functions (11, (II)) are used.The scanning line (
Assuming that the time on the line) is 1, the prediction error by prediction method (1) for the pixel time 13+1:2+i-] on the currently considered line is 5.4.3 as shown in Figure 1A. Assume that the prediction error by the prediction function (II) is 3.5.4 as shown in FIG. 1B. At this time, the pixel time of the line i −3, i
-2, i-1 is the one with the smaller prediction error due to both prediction functions (1) and (II), so the second T as shown in the row of the current line in the figure.
It becomes I, I, I.

Since the optimal prediction function at the previous pixel time i-1 is (1), if only the determination result at the previous pixel time is used, the prediction function X at the current pixel time (i) will be (1). . If the occurrence state of the optimal prediction function at a pixel time one pixel time or more before, one line before, and two lines before is shown in Fig. 2, then (1') in the vertical direction of the screen at pixel time i. , If we look at Bakun consisting of (TI), we can immediately see that (II) is more likely.

Such a phenomenon in which the optimal prediction function suddenly changes usually occurs at the contour part. Since the contour part is, so to speak, a part where the brightness is discontinuous, it is necessary to estimate the optimal prediction function in relation to the surroundings. Furthermore, since images usually contain a large number of contours, it is particularly important to utilize the surrounding relationships.

In other words, this invention improves the accuracy of estimation by estimating the optimal prediction function at the current pixel time by considering the spatial distribution state of the optimal prediction function corresponding to the encoded pixel (referred to as a pixel). Improving predictive coding efficiency,
In other words, the purpose is to reduce the amount of information to be transmitted or recorded.

In the present invention, on the transmitting side, one prediction signal estimated to be optimal is adaptively selected for each pixel from among prediction signals generated by a plurality of prediction functions, and the selected prediction signal is used. This is an adaptive predictive coding/decoding method in which a multilevel image signal is predictively encoded using a multilevel image signal, and a multilevel -1 image signal is predictively decoded on the receiving side using the opposite principle to that on the transmitting side. A plurality of encoded pixels (N pixels) in the vicinity of the pixel (current pixel) to be encoded are used as reference pixels for the current pixel, and which of the M prediction functions is the optimal prediction function for each reference pixel. Detect whether it was based on the prediction error,
゛When constructing MN reference pixel states from the detection results for N pixels and determining a prediction signal to be used for the current pixel from these, each of the MN states and the prediction function for the current pixel are set with high probability in each state. By selecting one prediction function that is the optimal prediction function, it is possible to make a one-to-one correspondence between the small number of prediction functions (at least up to the current time), and to
The method is characterized in that when a state appears, a prediction function to be used for the current pixel is determined according to this one-to-one correspondence.

Next, the third section describes a method for estimating the optimal prediction function for the current pixel from the generation state of the optimal prediction function for the reference pixel.
This will be explained in more detail with reference to FIG. FIG. 3 is a diagram showing an example of the arrangement of reference pixels, where X is the current pixel, A,
B and C indicate the positions of reference pixels. In the following, two types of prediction methods are used (11, (If)), and the optimal prediction 'tAll for the current pixel This will be explained as estimating a function.

When the number of reference pixels is 3, ta#, Figure 4 (1) ~
Eight states as shown in (8) can occur. Fourth
In the figure, "0" indicates that the prediction function (1) is the optimal prediction function, and "1" indicates that the prediction function (n) is the optimal prediction function for each reference pixel. In the present invention, in each state, the probability that the prediction function (Q, (II) will be the optimal prediction function for the current pixel (that is, the prediction error will be smaller than the other) is determined in advance. In other words, in each state, the number of pixels for which the prediction functions (1) and (II) are the optimal prediction function for the current pixel is individually counted (.

In the following, by such measurements, the probability that the prediction function (1) becomes the optimal prediction function in each state is determined by, for example, the fourth
The following explanation assumes that the value shown in the figure is P.

In the present invention, the frame, P□ becomes large (for example, pi≧0
．． 5) state of the reference pixel ((1) in the example in Figure 4),
In (2) and (5)), the prediction function (Il) is estimated to be the optimal prediction function for the current pixel, and in other cases, the prediction function (n) is estimated to be the optimal prediction function for the current pixel, and is used as the prediction function for the current pixel. If the preliminary tracing function applied to pixels is estimated in this way, the optimal prediction function can be estimated with high probability.

For example, in the example shown in Fig. 4, Fig. 4 (1), (
In states 2) and (5), the prediction function (
If summer) is used, the probabilities of correctly selecting the optimal prediction function are 0.97 and 0.7510.80, respectively.

Furthermore, when using prediction function (1) for the current pixel in each state of Figure 4 (3), (4), (6), (7), and (8), the optimal prediction function can be selected correctly. The probabilities are 0.80+0.75, 0.55, 0.85, and 0.98, respectively. Assuming that states (1) to (8) occur with equal probability, this method will yield (0,97+0.75+0.
80+0.80+0.75+0.55+0.85 +0
．． The optimal prediction function can be correctly selected with a high probability of 98 and 806, that is, 80% or more. In this way, predictive encoding is performed on the multivalued image signal with high efficiency.

Embodiments of the predictive encoding device according to the present invention will be described in detail below with reference to the drawings. FIG. 5 shows an example in which two prediction methods are used as the simplest case. A multivalued image signal inputted from an input terminal 11 is supplied to a subtracter 12 . In the subtracter 12, the predictor 1
The prediction signal selected by the selector 14 from among the two types of prediction signals generated in step 3 is received, and is compared with the input multivalued image signal supplied from the terminal 11. This difference, that is, the prediction error, is normally converted into a 1 by a digitizer 15 which limits the usage level in order to reduce the occurrence information, and then is supplied to a 1 code converter 16 and an adder 17. Adder 17
, a locally decoded signal is generated by adding the prediction signal supplied from the selector 14 and the quantized prediction error signal supplied from the quantizer 15. This locally decoded signal 45 is supplied to the predictor 13, and the predictor 13 generates a plurality of predicted signals using this locally decoded signal.

Two types of prediction σ (+10,000 expression) performed by the predictor 13 are 1!1 prediction (D
PCM) method and an inter-frame prediction method that predicts using the locally decoded 4G code of one frame before.This predictor 13 is a circuit in which a 1-pixel delay element and a 1-frame delay element are arranged in parallel. This will be realized in

Next, a method of generating a control signal for the selector 14 will be explained. The prediction signals generated by the predictor 13 using the two types of prediction methods are supplied to the selector 14 and at the same time are also supplied to the optimal prediction function determining circuit 18 . This decision circuit 1
8, which of the two types of prediction signals is sent to the adder 17?
It compares and determines whether the local decoded value of the multivalued image signal supplied from In this example, a 1-bit signal) is transferred to the storage circuit 19. For example, if the prediction method (1) described above is intra-frame prediction and the present prediction method (II) is inter-frame prediction,
If the predicted signal by intra-frame prediction is closer to the locally decoded signal value, "0" is output, otherwise "1" is output. The memory circuit 19 is composed of a memory and an element having a memory capacity of, for example, one line (one line in the case of a television signal), and the example shown in FIG. 3 is used at the current pixel time. Then, the information indicating the optimal prediction function to be applied to the previous pixel A1 and the pixels B and C of the previous line is the signal acid 5.
2 in parallel (3 bit parallel in this case), The signal is supplied to the circuit 21. This determination circuit 21 includes a storage circuit 19
Based on the signal indicating the occurrence status of the optimal prediction function at the reference pixel, which is the output of A signal that controls the selection operation of the selector 14 (for example, the signal value is "0" if it is determined that intra-frame prediction should be used; the signal value is 1 if it is determined that inter-frame prediction should be used).
g value "l") is supplied to the selector 14. - The predicted g4 difference iFt signal, which has been quantized by the child generator 15, is input to the code converter 16, encoded, for example, variable length encoded, and supplied to a transmission path or a recording medium.

Next, the determination circuit 21 will be explained in detail.

This determination circuit 21 is normally a read-only memory (Read
0nly Memory v ROM). In other words, in the above example, a 3-bit signal is supplied to the input address of the ROM, but the results of statistical judgments on various images are pre-assigned to the address corresponding to the 3-bit bit pattern at this time. For example, let us assume that the statistical judgment result is as shown in FIG. Assuming that a signal indicating the optimal prediction function for the reference pixel C is input to the upper bits, for example, the fourth
The input address corresponding to the reference pixel state shown in the figure is r ll0J, in which case PI<0
．． 5, so in this reference pixel state, the prediction function (1, i.e., interframe prediction) is applied to the current pixel.
The value written to OM is "l". The values written to other input addresses are determined in the same way, and in the example shown in Figure 4, the values shown in the output column of Figure 7 are for each input address of the ROM. It will be written. The determination circuit 21 is configured by the table written in the ROM in this manner.

Next, a decoding device for decoding information encoded in this manner will be described in detail with reference to FIG. The encoded multi-level image signal transmitted or read from the recording medium is supplied to the code inverse converter 23 through the terminal 22. In the code inverse converter 23, a prediction error signal is extracted from the encoded multi-I 1-Image 1-BO signal, and the unequal length code is inversely converted into the equal length code, thereby performing the following decoding.

The prediction error signal inversely converted into an equal length code by the code inverse converter 23 is supplied to the adder 24. In the adder 24,
Among the two prediction signals generated by the predictor 25, the selector 26
Upon receiving the prediction signal selected by , the prediction error signal and the output of the code inverse converter 23 are added to obtain a decoded multilevel m-image signal. This decoded multivalued image signal is simultaneously supplied to an external image signal output terminal 27, a predictor 25, and an optimal prediction method determining circuit 28, respectively.

Using this decoded image signal, the predictor 25 generates a plurality of predicted signals. The optimal prediction method determination circuit 28 uses this decoded image signal and a plurality of prediction signals (two types in this example) from the predictor 25 to determine the optimal prediction method determination circuit 18 of the encoding device described with reference to FIG. The optimal prediction method is determined according to the same rules as in the case, and the result is transferred to the storage circuit 29.

The memory circuit 29 is the memory circuit 1 in the encoding device of FIG.
Similarly to 9, in this example, 3-bit information indicating the optimal prediction method for 3 pixel times is supplied to the determination circuit 31. The determination circuit 31 controls the selection of the selector 26 using the same rules as the determination circuit 21 of the encoding device. Note that the predictor 25, selector 26, optimal prediction method determination circuit 28, storage circuit 29, and determination circuit 31 have the same configuration as the predictor 13, selector 14, optimal prediction method determination circuit 18, and storage in the encoding device, respectively. The circuit 19 and the determination circuit 21 are the same, and their mutual connection relationship is also the same.

As explained in detail above, according to the present invention, the optimal prediction function for the current pixel can be estimated with high probability from the occurrence situation of the optimal prediction function in the reference pixel, and the non-Kvca and multilevel image signals can be Further, in the present invention, the embodiments described below h1 can also be realized.

In the above explanation, the table for estimating the optimal prediction function for one pixel was determined through experiments using previously obtained measurement results as shown in FIG. 4, for example. In other words, the tape formula for estimating the optimal prediction function at the current pixel was fixed, but it is not necessary to fix the tape γ (see Figure 4) while performing encoding and decoding. For each of the reference pixel patterns in the reference pixels shown in (1) to (8), while measuring the probability of occurrence that the inter-frame prediction d11 and intra-frame prediction have the optimal F@ number, the optimal prediction function is estimated. It is also possible to update the table (this is called learning + 12 functions). An example in this case will be described below. Note that the example already explained in this example (first example)
The only different parts are the determination circuits 21 and 31, and since the determination circuits 21 and 31 have exactly the same configuration and function, only the determination circuit 21 will be described. The configuration of the determination circuit 21 used in this embodiment is shown in FIG.

In FIG. 8, a parallel 3-bit signal indicating the reference pixel state, which is the output of the unique circuit 19 shown in FIG. The table memory 65 is constituted by a random access memory, and its input address and output are all (same) as the ROM constituting the determination circuit 21 of the first embodiment.

The table memory 65 stores a table for estimating the optimal prediction function for the current pixel, which has been updated by the time before the input of the signal indicating the reference pixel state for the current pixel. A signal input through signal line 52 indicating the reference pixel state is supplied to an input address of table memory 65, and table memory 65 outputs a signal indicating the prediction function to be used for the current pixel onto signal line 53. In other words, when using interframe prediction.

If LJ111 is used, outputs "0". Histogram counter 6
0 indicates each of the reference pixel states (in this case, as shown in FIG.
1) to (8)), count the number of pixels for which interframe prediction and intraframe prediction are the optimal prediction functions, and calculate the number of pixels for which interframe prediction and intraframe prediction are the optimal prediction functions for each reference pixel state (. A histogram is created to determine which one has a higher probability of becoming the optimal prediction function.
An address signal indicating one of the reference pixel states is supplied from the signal line 67, and after counting the added time units, the reference l specified by this address signal is
I! In the II system state), a signal indicating the frequency at which the inter-frame prediction becomes the optimal prediction function is sent to the signal line 61, and a signal indicating the frequency at which the intra-frame prediction becomes the optimum prediction function is sent to the signal line 62.
Output to. The signals output to the signal lines 61 and 62 are input to the comparator 63. The comparator 63 compares both input signal values, and if the signal on the signal line 61 is larger (the frequency at which the interframe prediction becomes the optimal prediction function in the reference pixel state specified by the address generator 66), the comparator 63 compares the two input signal values. signal 110rxJ (corresponding to high
” is stored in the table memory 65 via the signal line 64.

The table reader 65 updates the table for estimating the optimal prediction function by writing the value input via the signal line 64 to the address supplied via the signal line 67. For example, the address supplied by this signal line 67 is set to 00 at every N frame (N>1) time as illustrated in FIG.
By sequentially changing the values from 0 to 111, the entire contents of the table memory 65 can be updated regularly. In addition, the random access memory that makes up the table memory
By performing both write and read operations independently, it is possible to update the team while performing the encoding operation. Also, immediately after updating the table, the histogram counter is set to 60.
The contents of are cleared, and creation of a histogram to be used at the next table update time is started.

Next, the configuration and operation of the histogram counter 60 will be explained with reference to FIGS. 9 and 10.

FIG. 9 is a block diagram for explaining the histogram counter 60, and FIG. 10 is a block diagram for explaining the histogram counter 60.
This is a block diagram for explaining the configuration and operation of the accumulator section, which is one of the components.

A signal indicating the state of the reference pixel input through the signal line 52 is input to the register 70.As described above, the signal line 52 is connected to the three vias indicating the optimal prediction function for the reference pixels A, B, C, IfC:i6. ) They are parallel signals, but in FIG. 9 they are shown as three 1-bit signals on signal lines 52-1 to 52-3.When the current pixel time is referenced, the signal line 5
2-1 is reference pixel C, 52-2 is reference pixel B, 52-3
The explanation will be made on the assumption that represents the optimal prediction function of the reference pixel A.

The signals on the signal lines 52-1 to 52-3 are delayed by one sample time by the register 70, and each signal &! It is output in parallel to 71-1 to 71-3. Therefore, signal line 71-1.71-2
．． At 71-3, signals indicating optimal prediction functions for pixels one pixel to the left of reference pixel C, one pixel to the left of reference pixel B, and one pixel to the left of reference pixel A appear, respectively. Further, as described above, a signal indicating the optimal prediction function for the reference pixel A appears on the signal line 52-3. In other words, the signal line 52-3 and the signal lines 71-1 to 71-3 show an optimal pre-suppression function at a pixel one sample before (previous pixel) and an optimal prediction function at a reference pixel for the previous pixel. Therefore, the aforementioned histogram can be created from the signals of the signal line 52-3 and the signal lines 71-1 to 71-3.

The signals m71-1 to m71-3 are input to the line decoder 72. The line decoder 72 sets 4 of any one of the signal lines 73-1 to 73-8 to 0, and the accumulator section (117
4 and accumulator (+1) 75. Here, the configurations of the storage parts (1) and (II) are completely the same. The accumulator section (1) 74 has the reference numeral 8, each connected to a signal M73-1 to M73-8 as shown in FIG.
It consists of 8 counters numbered 1 to 88. Further, the signal m5z-3 is input to the accumulator (1), and the signal line 52-3 is input to the accumulator (111).
A signal obtained by inverting the signal No. 3 by an inverter 77 is input. These signals are sent to the accumulator section (11, (
[Used as a select signal for [1]. In other words, when the signal on the signal line 52-3 is "1" (when the optimal prediction function of the previous pixel is interframe prediction), the 7-chim rake section (only 11 operates, and when it is "0") 1 =
[Only the Yumulator stand ++(n) operates. For example, if the signal on the signal line 52-3 is 11'', the accumulators are built in the accumulators L to fI+, and the signal lines 73-1 to 8
Among the 8 counters to which signals are supplied, only the counter to which 0 is supplied as an input signal is counted and amplified by 1. 7Kimulator part (IIl also, signal &!76
A similar operation is performed when the signal #52-3 is "1" (in this case, the signal #52-3 is 0). In this way, for each of the reference l111i search states, the number of picture constants for which the interframe prediction is the optimal prediction function is 7.
4, and the number of pixels for which the intra-frame prediction has become the optimal prediction function is counted in the 7-cure unit (n).

In this way, a histogram is created for determining which intra-frame prediction has a higher probability of becoming the optimal prediction 6111 function for each reference pixel state.

The count results of the 8 built-in counters are each signal I%! 77-1 to 8 and input to the selector 79, and the count results of the eight counters built in the accumulator section (■) 75 are read to the signal lines 78-1 to 78-8 and input to the selector 8o. is input.

The selector 79.8o selects one of the eight outputs according to the address signal inputted through the signal line 67, and outputs it to each signal output 61.62. That is, the frequency at which the inter-frame prediction becomes the optimum prediction function for the specified reference pixel state is output to the signal line 61, and the frequency at which the intra-frame prediction becomes the optimum prediction function is output to the signal #62.

In the manner described above, an optimal prediction function determination circuit having a learning function can be realized.

According to this second embodiment, the optimal prediction (1111 function) can be estimated by following the property of image clarity that changes from time to time, and more efficient encoding can be performed. In the explanation, an example V Kotori has been explained in which the predicted J lead number is adaptively changed to 2, but the present invention can be similarly applied to the case where 3 or more prediction functions are used. Similarly, it is also possible to use m-hundred pixels that are temporally close to the current pixel.

[Brief explanation of the drawing]

Figure 1 is a diagram showing an example of prediction error by the two-plant prediction method.
Figure 2 is a diagram showing an example of the spatial distribution of the optimal prediction function, Figure 3 is a diagram showing seven examples of the arrangement of reference pixels, and Figure 4 is a diagram showing an example of the spatial distribution of the optimal prediction function.
FIG. 5 is a block diagram showing an embodiment of the predictive coding system according to the present invention. 6 is a block diagram showing an embodiment of a post-coding device for decoding information encoded by the encoding device according to the invention of Bako 4. In FIGS. 5 and 6, 11, 22: Input terminal, 12: d calculator, 13.25
: Predictor, 14.26: Selector, 16: Code converter, 1
7.24: Adder, 18.28: Il suitable prediction method determining circuit, 23: Sign inverse converter, 27: Output terminal. FIG. 7 is a diagram showing an example of input/output logic of the determination circuit 21.31. FIGS. 8, 9, and 10 are block diagrams for explaining the configuration and operation of the determination circuit 21, 31 used in other embodiments of the present invention. 〆1 agent valve 11jT, i' 9. ,L,,! ',
+ □') Bu゛, 2 Fig. 1 82 Drawing element time?) i-3i-2t-1i-j+1 Fig. 1 (2> (3) (4) Gate-1 Buttocks=0.97 F=0 .75 P=0.20.
P=0.251 1
11ge I P=0.80 F,=0.45? =0.75
P=0.02 Fig. 5 C Doji 6 Fig. 7

Claims (1)

[Claims] 1. On the transmitting side, one prediction signal estimated to be optimal is adaptively selected for each pixel from among prediction signals generated by a plurality of prediction functions. An adaptive predictive encoding/decoding method for a multi-value image signal, in which a multi-value image signal is predictively coded using a prediction signal, and the multi-value image signal is predictively decoded on the receiving side using logic opposite to that on the transmitting side, the method comprising: A plurality of encoded pixels (N) in the vicinity of the pixel to be encoded at the current time (current pixel) are used as reference pixels for the current pixel, and the M prediction functions are applied to each of the reference pixels. Among them, which one is the optimal prediction function is detected based on the prediction error, and MN reference pixel states are constructed from the detection results for N pixels. From this, when determining the prediction signal to be used for the current 11 pixels, For each state and the prediction function for the current pixel, by selecting one prediction function that is the optimal prediction function with a high probability in each state (at least one-to-one correspondence up to the current time), the reference pixel An adaptive predictive encoding/decoding method for multi-level image signals whose characteristic is to determine a prediction function to be used for the current pixel according to this one-to-one correspondence when any l state among the states appears. 2. ) prediction signals generated by the prediction functions, one prediction signal estimated to be optimal is adaptively selected for each pixel, and the selected prediction signal is used to perform predictive coding of the multi-level image signal. An apparatus for adaptive predictive coding of a multivalued image signal, comprising: means for generating M prediction signals; means for selecting one prediction signal f from the M prediction signals; means for encoding a prediction error signal between the selected prediction signal and the multi-value image signal; means for locally decoding the multi-value image signal from the prediction error signal and the selected prediction signal; and the multi-value image signal. A prediction function (J
means for generating a signal indicative of a suitable prediction function 9; and a plurality of encoded signals (N) located in the vicinity of a pixel (current pixel) that is the output of the generating means and is encoded at the current time.
pixel as a reference pixel, a means for generating a signal group formed by signals indicating the optimal prediction function for each of the reference pixels and capable of indicating MN states (reference pixel states); each and a prediction function that is estimated to be optimal with a high probability for the current pixel at least once by the current time.
determining a pre<I11 function to be used for the current pixel from the one-to-one correspondence and the signal group indicating the reference pixel state, and selecting the prediction signal generated by the determined prediction function; 1. An adaptive predictive coding device for a multivalued image signal, comprising means for controlling the selection by the means. 3. A multivalued image that adaptively selects one prediction signal for each pixel from among the prediction signals generated by a plurality of prediction functions, and performs adaptive predictive coding using the selected prediction signal. An adaptive predictive decoding device that predictively decodes a signal, comprising M
means for generating M predicted signals; means for selecting one predicted signal from the M predicted signals; means for decoding a multi-value image signal from the predistributed distortion value image signal; and a prediction function such that the difference between the decoded multi-value image signal and the M pre-6III signals for the predistributed distortion value image signal is lj. (optimal prediction function);
A plurality of (N) pixel operations adjacent to the decoding ratio located in the vicinity of the pixel (current pixel) which is the output of the generating means and is decoded at the current time are used as reference pixels, and an optimal prediction function for each of the reference pixels is calculated. means for generating a group of signals formed by signals that overlap the current pixel and representing MN states (reference frame v, state) 1, each of the N11 states being estimated to be optimal with a high probability for the current pixel; a means for associating a prediction function one-to-one (at least up to the current time); and determining a prediction function to be used for the current pixel from this one-to-one correspondence and the signal group indicating the reference pixel state; an adaptive predictive decoding device for a multivalued image 16, comprising means for causing the selecting means to select a predicted signal generated by the above.