JPH0366861B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to an image signal encoding device that compresses and encodes an image signal with high efficiency. Typical of conventional image signal encoding devices
There is a DPCM encoding device. This uses the input image signal S (hereinafter referred to as the reference image signal) to determine the prediction signal x^ of the currently input image signal x (hereinafter referred to as the current image signal), and the prediction error signal e is determined by the difference between x and x^, e=x−x^ (1), and compression encoding is performed by taking advantage of the fact that the prediction error signal e tends to concentrate at small values around 0. . A case where, for example, three pixels a, b, and c in FIG. 1 are selected as the reference image signal S for the current image signal x will be explained in more detail. In FIG. 1, the dotted line indicates the main scanning line, along which the image is scanned from left to right. When the right end is reached, scanning continues from the left end along the main scanning line one line below. Image signals are input to the encoding device in scanning order. The predicted signal x^ is expressed as S=(a, b,
There are various ways to determine c), but pixels _{a ,} b, _c belonging to the reference image signal S are
It is often determined by a linear combination of k ₁ , k ₂ , and k ₃ are constants called prediction coefficients, and are determined so that the prediction error signal e concentrates on small values around 0. For example, if the root mean square value of e <e ² > _AV = <x-x^) ² > _AV = < (x-k ₁ a-k ₂ b-k ₃ c) ² > _AV (3) is stipulated by. The prediction error signal e determined in this way is 0 as shown in FIG.
Since it follows the centralized probability distribution P(e), it is compressed and encoded using this. FIG. 3 shows the code length L(e) for the prediction error signal e in conventional DPCM encoding. As shown in FIG. 3, a short code is assigned to a prediction error signal e having a small signal value with a high probability of occurrence, and a long code is assigned to a prediction error signal e having a large signal value having a low probability of occurrence.
When considered in more detail, the occurrence probability distribution of the prediction error signal e is considered to be different depending on the reference image signal S, and to be an occurrence probability distribution of P(e|S) for each S. Here, P(e|S) is the occurrence probability distribution of e under condition S, and is a different distribution for each S. P(e) in FIG. 2 described above is the average of P(e|S) with respect to S, P(e)= 〓 ^S ΣP(e|S) P(S) (4). Here, P(S) is the probability of occurrence of the reference image signal S, and Σ ^S means the sum of all S's. Equation (4) expresses P(e|S) for each S as S
It shows that P(e) is weighted and averaged by the probability of occurrence of . Give a concrete example, P(e|S)
Let me explain the difference between and P(e). The reference image signal S consists of three pixels (a, b, c) in FIG. 1, and the image signal has 16 values from 0 to 15 (4 bits/pixel). Suppose that the prediction coefficients k ₁ , k ₂ , and k ₃ are chosen to be 0.7, 0.2, and 0.1. When S is S ₁ =(8,8,8) (5), the predicted signal is x^=0.7×8+0.2×8+0.1×8=8 (6). When S=S ₁ , all three pixels a, b, c around the current image signal x are 8, and x is considered to be in the flat part of the image, and x is 8 and 7,
Very often it's 6. Therefore, when S=S ₁ , the occurrence probability distribution P(e|S ₁ ) of the prediction error signal e becomes a sharp distribution that almost concentrates on 0 and ±1 as shown in FIG. On the other hand, when S is S ₂ = (8, 11, 5) (7), the predicted signal is x^ = 0.7 x 8 + 0.2 x 11 + 0.1 x 5 = 8 (8). However, here the calculation results were rounded off to the nearest whole decimal point. When S=S ₂ , 3 pixels around x a,
The signal values of b and c vary greatly as 11, 8, and 5, and x is considered to be at the edge of the image. At this time, it is thought that x often takes a signal value between 11 and 5 (e is +3 to -3), but there is no signal value that occurs with a particularly high probability. Therefore, when S=S ₂ , the occurrence probability distribution P(e|S ₂ ) of the prediction error signal e becomes a wide and gentle distribution as shown in FIG. In this way, the type of P(e|S) differs greatly depending on S. P(e｜S) as P
The so-called average P(e|S) obtained by weighting the average by (s) is P(e) as shown in FIG.
Therefore, P(e) has a distribution that is gentler than P(e|S ₁ ) but sharper than P(e|S ₂ ). Traditional
In DPCM encoding, e is encoded on the assumption that the occurrence probability distribution of e is P(e), and the fact that the occurrence probability distribution of e changes depending on S as described above is not taken into consideration at all. Therefore, depending on S, the code length L(e) given in FIG. 3 is often not appropriate, resulting in a disadvantage of low compression ratio. An object of the present invention is to provide an image signal encoding device that eliminates such drawbacks and has a high compression rate. According to the present invention, a prediction error signal of an input multilevel image signal is divided into a plurality of groups for each pixel, and each pixel is unequal-length coded using a different unequal-length code for each group. An image signal encoding device that uses an image signal s that has already been input to generate a predicted signal x^ of a multilevel image signal x that is currently being input.
means for generating a prediction error signal e of the multivalued image signal x; means for generating a mode signal M indicating the degree of accuracy of the prediction using the degree of change in the image signal s; The prediction error signal e is divided into groups according to the signal value of the mode signal M, and for each group, a set of unequal length codes determined according to the probability distribution of the prediction error signal e in each group is used. An image signal encoding device characterized by having means for unequal length encoding the prediction error signal e is obtained. By using the present invention, image signals can be efficiently encoded. The present invention will be explained in detail below using the drawings. The image signal has 16 values from 0 to 15 (4 bits/pixel),
An example in which the reference image signal S consists of three pixels a, b, and c in FIG. 1 will be explained. First, a first example will be described. FIG. 6 is an example of a block diagram of the first embodiment of the encoding device of the present invention. In FIG. 6, an image signal (4 bits/pixel) is sequentially input to a 4(l+2) bit shift register 1. Here, l is the number of pixels per main scanning line, and the shift register 1 stores image signals of 1 line + 2 pixels. The pixel that has just been input to shift register 1 is x, the one that was input just before is a, the one that was input (l-1) pixels before is c, and the one that was input one line ago is c. It is b. Prediction signal generator 2
is the reference image signal S=(a,
b, c), and input a, c) as shown in equation (2).
A predicted signal x^ of x is created and output by linear calculation of b and c. The subtracter 3 extracts x^ from the prediction signal generator 2,
Input x from shift register 1, calculate the difference signal (see equation (1)), and output. This is the prediction error signal e of x. The mode signal generator 4 generates a reference image signal S=(a, b, c) from the shift register 1.
is input, and a mode signal M is output. The mode signal M takes two values, 0 and 1, and M=0 means that the prediction error signal e is so high that it concentrates on 0.
Conversely, 1 indicates a low degree of concentration at 0. As already explained in detail, the occurrence probability distribution P(e|S) of e varies greatly depending on S. Therefore, we type those in which P(e|S) is sharply concentrated at 0, such as S=S ₁ , and conversely, P(e|S), such as S=S _2.
S) that exhibits a gentle mountain shape is called a type, and S is classified into two types: type and type. That is, the type has a high zero concentration of P(e|S), and the type has a low concentration. The mode signal generator 4 generates M=0 for type S and M=1 for type S according to this classification, and the prediction error signal encoding circuit 5
Add to. The prediction error signal encoding circuit 5 operates differently depending on the signal value of M. That is, when M=0, the prediction error signal encoding circuit 5 generates a code that matches the prediction error signal e when S is a type (hereinafter referred to as type e), and when M=1, when S is a type. The prediction error signal e when
(hereinafter referred to as type e). The probability distribution of occurrence of type e is P
Using (e|S) and the probability of occurrence of S P(S), it is given by P〓(e)=〓 ^SI P(e|S)P(S) (9). Here, 〓 ^SI means the sum of types S. P〓(e) is P(e｜S ₁ )
This is a distribution that sharply concentrates on 0, as shown in Figure 7, because it is an average of the types of distributions that concentrate sharply on 0. When M=0, the prediction error signal encoding circuit 5 uses the fact that e follows a distribution P〓(e) that is sharply concentrated at 0 to convert it into a code F〓(e).
Compression encoding is performed using Figure 8 shows the length L〓(e) of the symbol F〓(e). L〓(e) is very short for 0, which has a very high probability of occurrence, and e that is close to it, and is made very long for e, which is far from 0, which has a very low probability of occurrence. On the other hand, the occurrence probability distribution of type e is calculated using P(e|S) and the occurrence probability distribution of S P(S): P〓(e)= 〓 ^SH P(e|S)P(S) (10) is given by Here, 〓 ^SH means the sum of types S. P〓(e) is P(e｜S ₂ )
As shown in Figure 9, the distribution is a gentle mountain-shaped distribution centered on 0, and is averaged. When M=1, the prediction error signal encoding circuit 5 takes advantage of the fact that the type e follows a gentle mountain-shaped distribution P〓(e) centered on 0, and converts it into a code F〓(e). and compression-encodes it. FIG. 10 shows the length L(e) of the symbol F(e). L〓(e) is relatively short for 0, which has a relatively high probability of occurrence, and e around it, and is relatively long for e, which is far from 0, which has a relatively low probability of occurrence. As can be seen by comparing Figures 8 and 10, L〓(e) is longer for e around 0 and shorter for e far from 0 than L〓(e). . The feature of the present invention is that the code suitable for each type and type e is as described above.
The purpose is to increase the efficiency of compression encoding by increasing the ratio of F〓(e) and F〓(e). In conventional DPCM encoding, the code length L(e) (see Figure 3) is L〓(e) (see Figure 8) and L〓
This is an intermediate type of (e) (see Figure 10). Code length L
Although (e) does not fit the probability distribution for either type e or type e, it is determined so that it does not deviate too much from both. On the other hand, in the present invention, type e
Since the code length is set to match the occurrence probability distribution of e, L〓(e) for type e, and L〓(e) for type e, it is more efficient than conventional DPCM encoding. In conventional DPCM encoding, the code length of e is determined based on the type and the distribution that is the average of the types. Therefore, in conventional DPCM encoding, when e is a type, the code for e near zero is longer than the type code, and the encoding efficiency decreases, even though the zero concentration of e is high. Also conventional
In DPCM encoding, when e is of type, e
Although the zero concentration of e is low, the coding efficiency decreases because the code for e that is far from zero is longer than the code of type e. In the end, the prediction error signal encoding circuit 5 converts the code F〓(e) when M=0 to
When M=1, select code F〓(e) and output code F(e)
shall be. FIGS. 11, 12, and 13 are block diagrams showing specific examples of the configuration of the predicted signal generator 2, respectively. The predicted signal generator shown in FIG. 11 is a linear predicted signal generator corresponding to equation (2). In Figure 11,
Prediction coefficients are applied to a, b, and c by multipliers 6, 7, and 8.
k ₁ , k ₂ , and k ₃ are multiplied, and the sum of the multiplication results is calculated by adder 9 and output as x^. The prediction signal generator shown in FIG. 12 changes the prediction coefficients k ₁ , k ₂ , k ₃ depending on the signal value of M. This is because the appropriate prediction coefficients k ₁ , k ₂ , k ₃ are considered to be different between a flat part of an image where M=0 and an edge part of an image where M=1. Using prediction coefficients that change in this way is expected to make predictions more accurate and improve coding efficiency. In FIG. 12, the selection circuits 10, 12, and 14 select prediction coefficients k ₁₁ , k ₂₁ , k ₃₁ when M=0, and select prediction coefficients k ₁₂ , k ₂₂ , k ₃₂ when M=1. Multiplier 11,1
3 and 15 perform multiplication of a, b, and c by the selected prediction coefficient. Adder 16 calculates the sum of these multiplication results and outputs it as x^. The predicted signal generator in FIG. 13 generates each reference image signal S=(a, b, c)
This is a method in which the predicted signal x^ is stored in the ROM 17. As mentioned above, by linear operation of a, b, c, x^
The method of determining is often used because it is simple, but the resulting x^ is not necessarily optimal when considered for each S. Let's consider the reasons below. Prediction coefficients k ₁ , k ₂ , k ₃
is determined, for example, using equation (3) so that the root mean square prediction error is minimized. Equation (3) can be considered as follows. <e ² > _AV = <(xx^) ² > _AV = Σ ^S Σ ^X P(x,S)(xx^) ² = Σ ^S Σ ^X P(S)P(x|S)(xx^) ² = Σ ^S P(S){ Σ ^X P(x|S)(xx^) ² } = Σ ^S P(S){ Σ ^X P(x|S)(xk ₁ ak ₂ bk ₃ c) ² } ( 11) Here, P(S) is the probability of occurrence of the reference image signal S,
P(x|S) is the conditional probability that the current image signal becomes x when the reference image signal is S. In equation (4), that is, equation (11), the mean square prediction error over all S is considered. By the way, the mean square prediction error < _es ² > _AV for each S is different from this, < _es ² > _AV = <(x _s − x^ _s ) ² > _AV = Σ ^X P(x|S) ( _{xx ^} ^{) 2} ₌ ^Σ The subscript S added to e, x, x^, k ₁ , k ₂ , k ₃ in equation (12) is especially the subscript S added to e, x, x^, k ₁ , k ₂ , k when the reference image signal is S. ₃ . ＜e _s
² > Set prediction coefficient for each S so that _AV is minimized
If k ₁₅ , k ₂₅ , and k ₃₅ are determined, the mean square prediction error for all S is given by <e ² > _AV = Σ ^S P(S) < e _s ² > _AV (13). Obviously, <e ² > _AV in (13) is smaller than <e ² > _AV in (11). However, in (11) the prediction coefficients k ₁ , k ₂ , k ₃ were constant regardless of S, but in (13) it is necessary to give different prediction coefficients k ₁ , k ₂ , k ₃ for each S. That is, it is necessary to store and prepare prediction coefficients k ₁ , k ₂ , k ₃ for each S. This ends up being each S.
This is the same as storing and preparing the predicted signal x^ for each time. In many cases, x^ _s that minimizes <e _s ² > _AV in (12) is the same as x that occurs with the highest probability in each S. FIG. 13 shows the configuration of a predicted signal generator using this method. In Figure 13,
The ROM 17 inputs the reference image signal S=(a, b, c) as an address signal, and outputs the contents of the memory at the corresponding address as a prediction signal x^. The image signal is 16 values (4 bits), the reference image signal S is 3 pixels a,
When consisting of B and C, the ROM 17 has a 12-bit address and a 4-bit output. The configuration of Figure 13 is
Although ROM is required, as explained in detail above, the prediction accuracy is higher than the configurations shown in FIGS. 11 and 12. FIG. 14, FIG. 15, and FIG. 16 are block diagrams showing the configuration of the mode signal generator 4, respectively.
There are various possibilities, but three specific examples are shown below. The mode signal generator in FIG. 14 calculates the degree of change in the image signal within the reference image signal S=(a, b, c) by Σ=(a-b) ² +(b-c) ² +(c -a) ² (14) Measured by the sum Σ of the square of the difference between the image signal and the image signal, M=0: Σ<θ (15) M=1: Σθ (16) When Σ is smaller than a certain threshold θ, M= 0, and when it is large, M=1. When the change in the image signal within S is small, it is considered to be a flat part of the image, and prediction is likely to be accurate, and P(e|S) often has a sharp type of distribution, so M=0. S
When the change in the image signal is large within the range, it is considered to be an edge part of the image, and prediction is difficult to be accurate.
Since S) often has a gentle mountain-type distribution, M=1. In FIG. 14, subtracters 18, 19, and 20 subtract a-b, b-c, and ca, and multiplier circuits 21, 22, and 23 subtract the respective squares (a-b) ² and (b -c) ² and (c-a) ² are calculated and added to the adder circuit 24. The adder circuit 24 calculates their sum Σ and adds it to the threshold circuit 25. The threshold circuit 25 has a total sum Σ as a threshold value θ
A mode signal M is output when M is smaller than M=0 and when it is larger than M=1. The mode signal generator in FIG. 15 calculates the degree of change in the image signal within the reference image signal S=(a, b, c) as D _MAX =Max {(ab) ² , (bc) ² , (ca) ² } (17) and the maximum value of the square of the difference between the image signal in the reference image signal
It is measured by D _MAX , M = 0: D _MAX < θ (18) M = 1: D _MAX θ (19) When D _MAX is smaller than a certain threshold θ, M = 0, and when it is larger, M = 1. . In FIG. 15, subtraction circuits 26, 27, 28 are a-b, b-c, c-
The multiplication circuits 29, 30, 31 calculate the respective squares (a-b) ² , (b-c) ² , (c-
a) Calculate ² and add it to the maximum value detection circuit 32. The maximum value detection circuit 32 detects (a-b) ² , (b-c) ² , (c
-a) Detect the maximum value _DMAX of ² and add it to the threshold circuit 33. The threshold circuit 33 outputs a mode signal of M=0 when D _MAX is smaller than the threshold θ, and M=1 when D MAX is larger. The mode signal generator in FIG. 16 calculates the accuracy of prediction for each reference image signal S=(a, b, c),
The difficulty of hitting is stored in the ROM and output as a mode signal M. When the prediction is likely to be correct, M=
0, when it is difficult to hit, M=1. In Figure 16
The ROM 34 inputs the reference image signal S=(a, b, c) as an address signal, and outputs the contents of the memory at the corresponding address as a mode signal M. When the image signal is 16 values (4 bits) and the reference image signal S consists of 3 pixels a, b, c, the ROM 34 has address 12.
bit, output 1 bit. The configuration of Figure 16 is
Although it requires ROM, it has the advantage of being able to arbitrarily set the standard for predictability. For example, the mean square prediction error <e ² _s > _AV = Σ ^e P(e|S) e ² (20) found for each reference image signal S can be used as the standard. Here, the subscript S of e _s particularly indicates e when the reference image signal is S. At this time, the mode signal M is as follows: M=0: <e ² _s > _AV <θ (21) M=1: <e ² _s > _AV θ (22) and <e ² _s > When _AV is smaller than a certain value θ Set it as 0, and set it as 1 when it is large. Further, the probability P(o|S) that the prediction error signal becomes 0 when the reference image signal is S can be used as a standard. At this time, mode signal M
M=0: P(o|S) θ (23) M=1: P(o|S) < (24) and P(o|S) is 0 when it is larger than a certain value θ, and 1 when it is smaller. stipulate. No matter what other criteria are used for predictability, it is only necessary to write in the ROM 34 the signal values of M for each S determined by the criteria. Of course, M can also be determined by comparing Σ, D _MAX and θ as in the configurations shown in FIGS. 14 and 15. Each S in each configuration
All you have to do is write M determined for each case in the ROM 34. FIG. 17 shows the prediction error signal encoding circuit 5 of FIG.
FIG. 2 is a block diagram showing an example. The image signal had 16 values from 0 to 15. In FIG. 17, the symbol ROM37
is the mode signal M and the prediction error signal e as the address signal
Input as A ₅ to _{A 0} , and when M=0 (A ₆ = 0), enter the sign F〓(e) for type e as M=1 (A ₆ =
1), the sign F〓(e) for type e is Q ₁₀
~ Output to Q ₀ . This is applied to parallel data input terminals D ₁₀ to _{D 0} of the shift register 39. The code length ROM 38 inputs the mode signal M and the prediction error signal e as address signals B ₆ to B ₀ , and M=0 (B ₆ =
0), the length L of code F〓(e)〓 is the two's complement of (e) M
=1 (B ₆ =1), the two's complement of the length L〓(e) of the code F〓(e) is output to P ₃ to P ₀ . This is counter 40
is added to the initial value setting terminals _C3 to _C0 . The shift register 39 has parallel data input terminals _D10 ~
The code F(e) input to D ₀ (F〓(e) when M=0, M
When =1, F〓(e)) is output one after another from the serial data output terminal SO in the order of D ₁₀ to D ₀ . At this time, the counter 40 calculates the length L(e) of the code F(e) (when M=0
L〓(e), when M=1, L〓(e)) is counted, and when the shift register 39 outputs a code of L(e) bits from SO, the shift operation is stopped there and the next code is started. Input data from the parallel data input terminal. Code ROM3 from Table 1 to Table 4
7. An example of the contents of the code length ROM is shown below. Tables 1 and 2 are for M=0 (A ₆ =0),
Tables 3 and 4 are for M=1 (A ₆ =1). The code ROM 37 has 11-bit output data terminals Q ₁₀ to Q ₀ . The number of bits at the output data terminal of is determined by the longest code length of F(e). When the length of code F(e) is shorter than the maximum code length of 11 bits,
Codes F(e) are stored in order from Q ₁₀ to Q ₀ direction. For example, in Table 1, when the ROM address is 2, the code ROM output data is 1101, which is
means Q ₁₀ = 1, Q ₉ = 1, Q ₈ = 0, Q ₇ = 1,
Q ₆ to Q ₀ are empty and can store anything. In the ROM address column, values are shown in decimal notation when _A5 to _A0 (same as _B5 to _B0 ) are considered as binary numbers with _A5 as MSB and _A0 as LSB. When the ROM address is 1, the prediction error signal e is a negative number, and A ₅
~A ₀ represents a negative number e in two's complement representation. For example, at ROM address -2, A ₅ to A ₀ is (111110) ₂
It is. The code length ROM output data field shows the code length.
Output data P ₃ to _{P 0} of the ROM 38 are described. P ₃ ~ _{P 0} is the code length L(e) (L〓(e) when M=0, when M=1
It is the two's complement of L〓(e). When taking two's complement, P ₃
~ Consider P ₀ as a binary number where P ₃ is the MSB (sign bit) and P ₀ is the LSB. P ₃ ~ in the code length ROM output data field
_P0 is considered to be a binary number without a sign bit, with _P3 being the MSB and _P0 being the LSB, and is expressed in decimal notation. For example 13
indicates P ₃ to P ₀ = (1101) ₂ . For reference, code length L(e) (M=
When M=0, L〓(e), when M=1, L〓(e)) is shown in decimal notation.

【table】

【table】

【table】

【table】

[Table] Figures 18 and 19 are the codes of Tables 1 to 4.
This is a code tree used to determine the contents of the ROM 37 and code length ROM 38. 18 and 19 are code trees for the code F〓(e) for M=0 and the code F〓(e) for M=1, respectively. At each branch point on the code tree, a code 1 is given when branching upward, and a code 0 is given when branching downward. There are 31 terminal ends of the code tree, each corresponding to a prediction error signal e (+15 to -15).
The corresponding signal value of e is written in <> at the end of the code tree, followed by the code F(e) (code F〓(e) for M=0 in Fig. 18, M
The code F〓(e)) for =1 is written. M=
Code tree for 0 (Figure 18), code table (Table 1,
Table 2) and the code tree for M=1 (Figure 19),
As can be seen by comparing the code tables (Tables 3 and 4), the code length of type (M=0) code F〓(e)
L〓(e) is the code length of code F〓(e) of type (M=1)
Compared to L〓(e), it is shorter for e around 0, and 0
It is set longer for e that is far from . The operation of the prediction error signal encoding circuit shown in FIG. 17 will be explained in detail using the time chart shown in FIG. 20. Each signal name written in the left corner of Fig. 20 is
This corresponds to the signal name given to each signal line in Figure 7.
However, since the corresponding signal lines for t, L(e), and COUT are not shown in FIG. 17, they will be explained. t is a time noted for the convenience of the following explanation, and is increased by 1 when the CLK signal generates one pulse. L(e) is the code length of code F(e), and P ₃ ~
P ₀ is the two's complement of L(e). COUT indicates the contents of the counter 40, and initial values are given by _P3 to _P0 . The operation of the prediction error signal encoding circuit shown in FIG. 17 will be explained below using the time chart shown in FIG. When the STAT signal is applied at time 0, the prediction error signal encoding circuit starts operating. first
The counter 40 is cleared by the STAT signal.
At time 1, the pulse generator 35 generates a pulse signal DSTA at the falling edge of the STAT signal and applies it to the OR circuit 36. The OR circuit 36 passes the DSTA signal and outputs it as an RQST signal. The shift register 1 in FIG. 6 performs a shift operation based on the CLK signal when the RQST signal is H, and outputs the first current image signal x.
Enter. Accordingly, the prediction signal generator 2 and mode signal generator 4 shown in FIG. 6 generate a prediction signal x^ and a mode signal M for the first current image signal x. However, it is assumed that the contents of shift register 1 are cleared to 0 before the first x is input. And the prediction error signal e for the first x
is calculated from x and x^ by the subtracter 3, and the code ROM 37 (code length ROM
38) as address signals A ₅ -A ₀ (B ₅ -B ₀ ). The mode signal M is applied to the code ROM 37 (code length ROM 38) as address signals A ₆ and B ₆ . In the case of the time chart in Figure 20, the first
The th e is 1 and M is 0. The code ROM 37 has address signals M=0, e=1 (A ₆ =0, A ₅ to _{A 0}
=1), the first code F(e) is outputted to Q ₁₀ to Q ₈ as (101) ₂ (see FIG. 1). Code length ROM3
8 is an address signal M=0, e=1 (B ₆ =0, B ₅ ~
B ₀ =1), the one's complement number 13 (see Table 1) of the code length L(e)=3 for the first code F(e) is output. This is set in the counter 40 as an initial value. This setting operation is performed by the counter 40.
When the RQST signal (H) is applied to the LD terminal, it is applied at the end of that time. Therefore, at time 2, the content COUT of the counter 40 changes from 0 to 13.
The shift register 39 outputs the code ROM 37 output Q ₁₀
Load ~Q ₀ via parallel data input terminals D ₁₀ ~ _{D 0} . This load operation is performed by shift register 3.
When the RQST signal (H) is applied to the LD terminal of 9, it is applied at the end of that time. Therefore, the output signal SO of the shift register 39 at time 2 is the same as that at time 1.
It becomes equal to Q ₁₀ =1. At times 3 and 4, the shift register 39 performs a shift operation using the CLK signal,
Outputs codes 0 and 1 sequentially from the SO terminal. At this time, the counter 40 performs a count-up operation based on the CLK signal, and its contents COUT change to 14 and 15. When the content of the 4-bit counter 40 becomes COUT=(15) ₁₀ at time 4, it generates a CARY signal and detects the completion of output of the first code F(e).
The CARY signal passes through the OR circuit 36, becomes RQST, and is added to the shift register 1 in FIG. When the RQST signal becomes H, the shift register 1 performs a shift operation and inputs the second current image signal x. Accordingly, the prediction signal generator 2 and mode signal generator 4 shown in FIG. 6 generate a prediction signal x^ and a mode signal M for the second current image signal x. Then, the prediction error signal e for the second x is calculated by the subtracter 3 and added to the prediction error signal encoding circuit shown in FIG. Thereafter, the present prediction error signal encoding circuit performs the same operation as for the first prediction error signal e. In the time chart of FIG. 20, the second prediction error signal e is 0.
At time 4, the code ROM 37 has the code F(e)=(0)
₂ (Q ₁₀ =0), and the code length ROM 38 outputs 1's complement (15) ₁₀ of L(e). Then, at time 5, the shift register 39 receives the sign F(e) from the SO terminal.
=(0) Outputs ₂ (length of 1 bit). Time 5
, the counter 40 is L from the code length ROM 38.
One's complement of (e) (15) Immediately after loading ₁₀
Generate CARY signal. CARY signal is OR circuit 3
6 and becomes the RQST signal, which requests the shift register 1 of FIG. 6 to perform a shift operation and input the third current image signal x. The prediction error signal shown in FIG. 17 repeats the same operation thereafter in accordance with the time chart shown in FIG. Note that the x mark on the time chart indicates that the signal value is indefinite or may be any value.
This concludes the description of the first embodiment. Next, a second embodiment of the present invention will be described.
FIG. 21 is a block diagram showing a second embodiment of the present invention. Below, as in the first embodiment, the image signal is 0.
The explanation will be made assuming that the reference image signal consists of the three pixels a, b, and c in FIG. 1. In FIG. 21, the subtracter 41 calculates the difference between the current image signal x (0 to 15) and its predicted signal x^ (0 to 15) (Equation (1)
Reference) Output as prediction error signal e (+15 to -15). The quantizer 42 generates a prediction error signal e (+15 to -
15) Quantize 5 bits and convert to signal y that can be expressed with 5 bits or less. The adder 45 adds the quantized prediction error signal y and the prediction signal x^ to produce a locally decoded signal.
Create x′. Local decoded signal x' is input to shift register 44. An example of shift register 44 is shown as shift register 48 in FIG. 48 is a 4(l+2) bit shift register. Here, l is the number of pixels per main scanning line. The local decoded signal just input to the shift register 48 is x', the local decoded signal input immediately before is a', the local decoded signal input (l-1) pixels before is c', l pixel before. That is, the locally decoded signal input one line before is b'. The predicted signal generator 43 receives the locally decoded signal from the shift register 44.
Input a′, b′, c′ and based on them, predict signal x^ of x
Output. The configuration of the predicted signal generator 43 in FIG. 21 is similar to that of the predicted signal generator 2 in FIG. 6 in the first embodiment (see FIGS. 11 to 13).
is possible. The mode signal generator 47 receives local decoded signals a', b', c' from the shift register 47 and outputs a mode signal M based thereon. The configuration of the mode signal generator 47 in FIG. 21 may be similar to that of the mode signal generator 4 in FIG. 6 in the first embodiment (see FIGS. 14 to 16). The prediction error signal encoding circuit 48 compresses and encodes the quantized prediction error signal y and outputs it as a code F(y).
In particular, the prediction error signal encoding circuit 48
=1, a different operation is performed. That is, M=0
When M=1, select the appropriate code F〓(y), and when M=1, select the appropriate code F〓(y), and select F(y).
Output as . The configuration of the prediction error signal encoding circuit 46 in FIG. 21 is similar to that of the prediction error signal encoding circuit 5 in FIG. 6 in the first embodiment (the first
(see Figure 7) is possible. FIG. 23 is a diagram showing an example of quantization characteristics of the prediction error signal quantization circuit 42. In FIG. 23, the horizontal axis shows the input prediction error signal e, and the vertical axis shows the output quantized prediction error signal y. In FIG. 23, a black circle at a break point of the quantization characteristic indicates that the quantization characteristic curve passes through that break point, and a white circle indicates that the quantization characteristic curve does not pass through that break point. According to this quantization characteristic, +15 to -15
e takes on 31 values: +15, +8, 0, -8, -
It is converted to y, which takes on 5 values of 15. y, which takes on 5 different values, can be expressed using 3 bits. This quantization circuit is a circuit that converts 5 bits e into y that can be expressed in 3 bits. The second embodiment differs from the first embodiment in that a prediction error signal quantization circuit 42 is present. In the second embodiment, the prediction error signal e is roughly quantized into y and then compressed and encoded, so that the encoding efficiency is better than in the first embodiment. However, the locally decoded signal x' differs from x due to quantization noise (the value of y slightly deviates from e depending on the quantization circuit. This refers to the amount of deviation). Since the decoding device paired with the encoding device of the second embodiment only obtains x' as a decoded signal, the image signal x is not completely decoded and only a slightly different image signal x' is obtained. There is. On the other hand, the decoding device paired with the encoding device of the first embodiment can obtain exactly the same decoded signal as the image signal x. This concludes the description of the second embodiment. Through the above first and second embodiments, the mode signal M
Although a binary signal is used, it is also possible to use a 3-value, 4-value, or more multi-value signal. Then, the prediction error signal e (quantized prediction error signal y) is 3
It can be encoded after classification, four or more classifications, and e classified into each type can be encoded.
If the code ratio is exactly matched to (y), the coding efficiency can be further improved.
Also, throughout the first and second embodiments, the reference image signal S
consists of 3 pixels a, b, c, but 4 pixels,
The number of pixels may be increased to 5 pixels or more to make the prediction more accurate. Furthermore, the image signal is always 16 from 0 to 15.
Although the explanation has been made using a value (4 bits), encoding devices for image signals having other numbers of levels can be constructed in a similar manner. Furthermore, in the first and second embodiments, the difference signal between the current image signal x and its predicted signal x^ (see equation (1)) was used as the prediction error signal e, but the prediction error signal It is also possible to use the ranking of facsimile signals (IEICE technical research report, CS79-176, Mizuno et al., "Coding of halftone facsimile signals"). As described above, according to the present invention, image signals can be encoded more efficiently than conventional DPCM encoding.

[Brief explanation of drawings]

Figure 1 is a diagram showing an example of the positional relationship between the image signal being input and the reference image signal, Figures 2, 4, 5, 7, and 9 are diagrams showing examples of various prediction error signal generation probability distributions, , 8 and 10 are diagrams showing examples of the lengths of various codes for prediction error signals. Figures 11, 12 and 13 are diagrams showing examples of various predicted signal generators, Figures 14, 15 and 16
The figure shows an example of various mode signal generators.
FIG. 7 is a diagram showing an example of a prediction error signal encoding circuit,
Figures 18 and 19 are diagrams showing examples of code trees;
22 is a diagram showing an example of a time chart of an encoding device, FIG. 22 is a diagram showing an example of a shift register,
FIG. 23 is a diagram showing an example of the quantization characteristics of the prediction error signal quantization circuit. In the figure, 1, 44, 48...shift register for storing image signals, 2, 43... prediction signal generator, 4, 47... mode signal generator, 3, 18,
19, 20, 26, 27, 28, 41... Subtraction circuit, 9, 16, 24, 45... Addition circuit, 5, 4
6...Prediction error signal encoding circuit, 6, 7, 8, 1
1, 13, 15, 21, 22, 23, 29, 3
0, 31... Multiplication circuit, 10, 12, 14... Selection circuit, 17... Prediction signal generation ROM, 25, 3
3... Threshold circuit, 32... Maximum value detection circuit, 34
...Mode signal generation ROM, 35...Pulse generation circuit, 36...OR circuit, 37...Sign ROM,
38... Code length ROM, 40... Counter, 39
. . . Shift register for code output, 42 . . . Prediction error signal quantization circuit.

Claims (1)

[Claims] 1. An image signal encoding device that divides the prediction error signal of an input multilevel image signal into a plurality of groups for each pixel, and performs unequal length encoding for each pixel using an unequal length code that differs for each group. And,
means for generating a prediction signal x^ of the currently input multi-value image signal x using the already input image signal s; means for generating a prediction error signal e of the multi-value image signal x; and the image signal s. means for generating a mode signal M indicating the degree of accuracy of the prediction using the degree of change in the prediction error signal e; dividing the prediction error signal e into groups according to the signal value of the mode signal M; An image signal encoding device comprising means for unequal-length encoding a prediction error signal e using a set of code length codes determined according to the occurrence probability distribution of the prediction error signal e.